very long comlillation time

[Mykola Matvyeyev]

1. Compilation time increases greatly when sequentially add lines to the cond form (see below).

I would like to get less compilation time.

(defun cond-bug (value)
  "Square brackets contain compilation time"
 (cond
   ((< value 0 ) :one) ; [0.02 secs]
   ((< 0 value 100 ) :two) ; [0.02 secs]
   ((< 100 value 200 ) :three) ; [0.03 secs]
   ((< 200 value 300 ) :four) ; [0.04 secs]
   ((< 300 value 400 ) :five) ; [0.06 secs]
   ((< 400 value 500 ) :six) ; [0.15 secs]
   ((< 500 value 600 ) :seven) ; [0.25 secs]
   ((< 600 value 700 ) :eight) ; [0.51 secs]
   ((< 700 value 800 ) :nine) ; [0.98 secs]
   ((< 800 value 900 ) :ten) ; [1.64 secs]
   ((< 900 value 1000 ) :eleven) ; [2.76 secs]
   ((< 1000 value 1100 ) :twelve) ; [4.35 secs]
   ((< 1100 value 1200 ) :thirteen) ; [6.75 secs]
   ((< 1200 value 1300 ) :fourteen) ; [9.95 secs]
   ((< 1300 value 1400 ) :fifteen) ; [14.38 secs]
   ((< 1400 value 1500 ) :sixteen) ; [19.89 secs]
   ((< 1500 value 1600 ) :seventeen); [29.36 secs]
   ((< 1600 value 1700 ) :eighteen) ; [41.66 secs]
   ((< 1700 value 1800 ) :nineteen) ; [53.28 secs]
   #+nil((< 1800 value 1900 ) :twenty) ; [65.26 secs]
   ))

2. $ sbcl --version
SBCL 2.2.9

3. $ uname -a
MINGW64_NT-10.0-22621 uakazi-note 3.3.6-341.x86_64 2022-09-20 22:07 UTC x86_64 Msys

4. *FEATURES*
(:QUICKLISP :ASDF3.3 :ASDF3.2 :ASDF3.1 :ASDF3 :ASDF2 :ASDF :OS-WINDOWS
 :NON-BASE-CHARS-EXIST-P :ASDF-UNICODE :X86-64 :GENCGC :64-BIT :ANSI-CL
 :COMMON-LISP :IEEE-FLOATING-POINT :LITTLE-ENDIAN :PACKAGE-LOCAL-NICKNAMES
 :SB-LDB :SB-PACKAGE-LOCKS :SB-SAFEPOINT :SB-THREAD :SB-UNICODE :SBCL :WIN32)

[Stas Boukarev (stassats)]

That's another instance of the constraint changes for join-type-constraints slowing things down.

[Christophe Rhodes (csr21-cantab)]

I appreciate that this might be a reduced test case, and that the real situation might not look like this -- and that also the behaviour is bad in any case. But just in case this helps make progress in the real case, note that the integer ranges are not contiguous -- the type of VALUE at the default branch of the COND includes every round hundred value. I found that

(defun cond-bug (value)
  "Square brackets contain compilation time"
 (cond
   ((< value 0 ) :one) ; [0.02 secs]
   ((< -1 value 100 ) :two) ; [0.02 secs]
   ((< 99 value 200 ) :three) ; [0.03 secs]
   ((< 199 value 300 ) :four) ; [0.04 secs]
   ((< 299 value 400 ) :five) ; [0.06 secs]
   ((< 399 value 500 ) :six) ; [0.15 secs]
   ((< 499 value 600 ) :seven) ; [0.25 secs]
   ((< 599 value 700 ) :eight) ; [0.51 secs]
   ((< 699 value 800 ) :nine) ; [0.98 secs]
   ((< 799 value 900 ) :ten) ; [1.64 secs]
   ((< 899 value 1000 ) :eleven) ; [2.76 secs]
   ((< 999 value 1100 ) :twelve) ; [4.35 secs]
   ((< 1099 value 1200 ) :thirteen) ; [6.75 secs]
   ((< 1199 value 1300 ) :fourteen) ; [9.95 secs]
   #+nil ((< 1300 value 1400 ) :fifteen) ; [14.38 secs]
   #+nil ((< 1400 value 1500 ) :sixteen) ; [19.89 secs]
   #+nil ((< 1500 value 1600 ) :seventeen); [29.36 secs]
   #+nil ((< 1600 value 1700 ) :eighteen) ; [41.66 secs]
   #+nil ((< 1700 value 1800 ) :nineteen) ; [53.28 secs]
   #+nil((< 1800 value 1900 ) :twenty) ; [65.26 secs]
   ))

compiled 50 times faster than the equivalent with x00 in the clauses

[Stas Boukarev (stassats)]

I pushed another optimization, it goes from 14 seconds to 3.5. Still too much for such a small form.

[Charles (karlosz)]

Now this form takes almost no time to compiler, so closing. (It must be some of the new type system changes that fixed this - it doesn't seem constraint related.)