fix #16434, superlinear compiler run time on large functions

[JeffBezanson]

Fixes this case completely, and seems to make the compiler a bit faster in general. Also helps #6685. Here are timings for these two issues on 0.4, master, and this branch:

0.4:
#16434:
     "n"    "t"         "t/n"        "t/n^2"  
   10.0    0.00164358  0.000164358  1.64358e-5
  100.0    0.00879924  8.79924e-5   8.79924e-7
 1000.0    0.473102    0.000473102  4.73102e-7
 2000.0    1.52983     0.000764916  3.82458e-7
 5000.0   11.0141      0.00220281   4.40563e-7
#6685:
  0.098373 seconds (287.72 k allocations: 9.650 MB)
  7.901539 seconds (23.98 M allocations: 1013.564 MB, 5.49% gc time)

master:
#16434:
     "n"   "t"         "t/n"        "t/n^2"  
   10.0   0.00416715  0.000416715  4.16715e-5
  100.0   0.0140538   0.000140538  1.40538e-6
 1000.0   0.508589    0.000508589  5.08589e-7
 2000.0   1.47        0.000734999  3.675e-7  
 5000.0   8.78618     0.00175724   3.51447e-7
#6685:
  0.115415 seconds (152.93 k allocations: 6.113 MB)
  6.339654 seconds (21.08 M allocations: 456.994 MB, 2.68% gc time)

this branch:
#16434:
     "n"   "t"         "t/n"        "t/n^2"  
   10.0   0.00379853  0.000379853  3.79853e-5
  100.0   0.0104883   0.000104883  1.04883e-6
 1000.0   0.139594    0.000139594  1.39594e-7
 2000.0   0.300802    0.000150401  7.52006e-8
 5000.0   0.723599    0.00014472   2.8944e-8 
#6685:
  0.109920 seconds (152.26 k allocations: 6.093 MB)
  5.517983 seconds (16.24 M allocations: 325.356 MB, 3.00% gc time)

[toivoh]

Thanks!