Device-side broadcast regression on 1.5

[marius311]

The following works with CUDA 1.2.1 and Julia 1.4 but errors on 1.5:

using CUDA

function foo(I,M)
    M[2I:2I+1] .= (1,2)
end

function bar(M)
    foo(1,M)
    nothing
end

M = CuVector{Cint}(undef, 4)
@cuda bar(M)

ERROR: LoadError: InvalidIRError: compiling kernel bar(CuDeviceArray{Int32,1,CUDA.AS.Global}) resulted in invalid LLVM IR
Reason: unsupported call through a literal pointer (call to jl_alloc_string)
Stacktrace:
 [1] _string_n at strings/string.jl:60
 [2] StringVector at iobuffer.jl:31
 [3] #IOBuffer#331 at iobuffer.jl:114
 [4] print_to_string at strings/io.jl:133
 [5] string at strings/io.jl:174
 [6] throwdm at broadcast.jl:1008
 [7] copyto! at broadcast.jl:922
 [8] copyto! at broadcast.jl:886
 [9] materialize! at broadcast.jl:848
 [10] materialize! at broadcast.jl:845
 [11] foo at Untitled-1:46
 [12] bar at Untitled-1:50

Manifest.toml

  [621f4979] AbstractFFTs v0.5.0
  [79e6a3ab] Adapt v2.0.2
  [b99e7846] BinaryProvider v0.5.10
  [fa961155] CEnum v0.4.1
  [052768ef] CUDA v1.2.1
  [864edb3b] DataStructures v0.17.20
  [e2ba6199] ExprTools v0.1.1
  [0c68f7d7] GPUArrays v5.0.0
  [61eb1bfa] GPUCompiler v0.5.5
  [929cbde3] LLVM v2.0.0
  [1914dd2f] MacroTools v0.5.5
  [872c559c] NNlib v0.7.4
  [bac558e1] OrderedCollections v1.3.0
  [189a3867] Reexport v0.2.0
  [ae029012] Requires v1.0.1
  [a759f4b9] TimerOutputs v0.5.6

[maleadt]

Device-side broadcast isn't really well-supported / tested.

[maleadt]

This looks like an inference change in the throwdm method, which we can't easily override or work around. Since device-side broadcast generally doesn't work, I don't feel like redefining the entire broadcast machinery to get rid of that specific error...

[marius311]

For posterity, a solution I found (mentioned on Slack) was to replace the broadcast

M[2I:2I+1] .= (1,2)

with

M[2I:2I+1] .= @SVector[1,2]

Understandable not to want to mess with it esp. since it has workarounds, but IMO its a pretty convenient feature to have, so if an easy path to get it back in the future presented itself, I think it'd be worth doing.