optimizers for LDB result in pessimal code

[Douglas Katzman]

(defun foo ()
 (let ((x (random (ash 1 128))))
 (loop repeat 1000 sum (loop for i below 128 by 8 sum (ldb (byte 8 i) x)))))

(defun bar ()
  (declare (notinline sb-kernel:%ldb)) ; add this line
  (let ((x (random (ash 1 128))))
   (loop repeat 1000 sum (loop for i below 128 by 8 sum (ldb (byte 8 i) x)))))

BAR is 50% faster and does not cons.

* (time(foo))
Evaluation took:
  0.001 seconds of real time
  0.001063 seconds of total run time (0.001063 user, 0.000000 system)
  100.00% CPU
  1,280,780 processor cycles
  360,432 bytes consed

* (time(bar))
Evaluation took:
  0.001 seconds of real time
  0.000000 seconds of total run time (0.000000 user, 0.000000 system)
  0.00% CPU
  774,976 processor cycles
  0 bytes consed

This is because LDB wants to become ASH and other stuff, which operate on bignums, but the full call to %LDB has a special case for producing a fixnum when possible. But if you declare LDB notinline then you're in worse shape because you cons a BYTE spec.