Work around #29952.

[maleadt]

Proposed hotfix for JuliaGPU/CUDAnative.jl#278, easy to backport. print_to_string is a very common function; even though in the case of GPU code it typically doesn't get emitted or optimized away, after #28876 it requires recursion which is GPU incompatible. Unless of course there's a easy fix for #29952.

[KristofferC]

Could we put in a comment here.

[KristofferC]

This adds an extra allocation to string

julia> @btime string(3, "f") # PR
  260.592 ns (6 allocations: 272 bytes)
"3f"

julia> @btime string(3, "f") # 1.0.1
  229.043 ns (5 allocations: 256 bytes)
"3f"

But I guess it is worth it :/

[JeffBezanson]

Ok, fine, but this seems pretty brittle. It's also not great that we have these recursion cycles everywhere because of string concatenation to throw errors in low-level routines. We will have to reckon with that somehow.

[maleadt]

Ok, fine, but this seems pretty brittle.

Yeah, this is just to have the upcoming 1.0.2 release GPU compatible. We'll have to find a better fix.

[maleadt]

Although fixing recursion might not be very hard. Observing it comes from print_to_string constructing a IOBuffer whose inner constructor can throw a ArgumentException with string interpolation (itself calling string->print_to_string again):

diff --git a/base/iobuffer.jl b/base/iobuffer.jl
index 7633581508..253f42e1d3 100644
--- a/base/iobuffer.jl
+++ b/base/iobuffer.jl
@@ -89,7 +89,7 @@ function IOBuffer(
         maxsize::Integer=typemax(Int),
         sizehint::Union{Integer,Nothing}=nothing)
     if maxsize < 0
-        throw(ArgumentError("negative maxsize: $(maxsize)"))
+        throw(ArgumentError("negative maxsize"))
     end
     if sizehint !== nothing
         sizehint!(data, sizehint)

Looks more "reasonable", at least. Also goes back to 5 allocations for calling string.