Reproduction scenario for nim --gc:arc performance difference

arcperfdiff.nim

import memfiles
import streams

proc mmapStream(mmFileName: string): Stream =
  return newMemMapFileStream(mmFileName)

iterator tokens(mmfile: Stream): string =
  var i,j = 0
  while not mmfile.atEnd():
    let c = mmfile.readChar()
    if c <= chr(32):
      if i < j:
        mmfile.setPosition(i)
        let s = mmfile.peekStr(j-i)
        yield s
        mmfile.setPosition(j+1)
      i = j+1
    inc j

when isMainModule:
  proc test(): void =
    var n = 0
    var l = 0
    let mmStream = mmapStream("set.mm")
    defer: mmstream.close()
    for t in tokens(mmStream):
      l += len(t)
      inc n
    echo(n," ",l)
  test()

run.sh

#!/bin/bash -xe
if [ ! -f set.mm ]
then
        curl --location https://github.com/metamath/set.mm/raw/b0925f0afd5963577ea76b252cb6613c885b393d/set.mm > set.mm
fi
nim c --opt:speed -d:release arcperfdiff && time ./arcperfdiff
nim c --gc:arc --opt:speed -d:release arcperfdiff && time ./arcperfdiff

# both should output
# 
# 4961591 29826787
# 
# but with nim 1.2.0 binaries from nixpkgs, for me
# the default version takes ~1 second,
# while the --gc:arc version takes 5+ seconds.

Compile with -d:danger instead of -d:release, there's still a ton of boundary checks in the inner loop causing slow downs.

Thanks! Is there a GitHub issue for this, or should I create one?