Re: [Sbcl-devel] problem with nested modular ash

[Alexey D. <ade...@co...>]

Hello,

Juho Snellman <js...@ik...> writes:

> (defun foo (x) 
>   (declare (optimize (speed 3) (safety 0)) (type (unsigned-byte 32) x)) 
>   ;; Swap the arguments to LOGAND around
>   (logand (ash (ash x 12) 0) #xffff)) ;; Can't inline ASH
...
> Does anyone have ideas on what's going on?

By the time of the last call of CUT-TO-WIDTH, the component looks
like this:

IR1 block 6 start c6
  6>     bind CLAMBDA FOO
successors c7

IR1 block 5 start c7
  7>     entry NIL
  8>  9: LOGAND {GLOBAL-FUNCTION}
 10> 11: CLAMBDA "LAMBDA-inlined ASH"
 12> 13: X
 14> 15: '12
 16> 17: '0
 18>     local combination v11 v13 v15 v17
successors c4

IR1 block NIL start c4
  4>     bind CLAMBDA "LAMBDA-inlined ASH" :KIND :LET
 19> 20: ASH {GLOBAL-FUNCTION}
 21>  2: VALUE
 22> 23: + {GLOBAL-FUNCTION}
 24> 25: AMOUNT1
 26> 27: AMOUNT2
 28> 29: cast v25 -[#<NAMED-TYPE *> -> #<NUMERIC-TYPE NUMBER>]
 30> 31: cast v27 -[#<NAMED-TYPE *> -> #<NUMERIC-TYPE NUMBER>]
 32>  3: known combination v23 v29 v31
 33> 34: cast v2 -[#<NAMED-TYPE *> -> #<NUMERIC-TYPE INTEGER>]
 35> 36: cast v3 -[#<NAMED-TYPE *> -> #<NUMERIC-TYPE INTEGER>]
 37>  1: known combination v20 v34 v36
 38>  5: '65535
 39> 40: known combination v9 v1 v5
successors c41

IR1 block 4 start c41
 41>     return v40 CLAMBDA FOO
successors c42

IR1 block 3 start c43
 43>     bind CLAMBDA "XEP for SB-C::FOO" :KIND :EXTERNAL
 44> 45: CLAMBDA FOO
 46> 47: #:G1
 48> 49: cast v47 +[#<NAMED-TYPE *> -> #<NUMERIC-TYPE (UNSIGNED-BYTE 32)>]
 50> 51: local combination v45 v49
successors c52

IR1 block 2 start c52
 52>     return v51 CLAMBDA "XEP for SB-C::FOO"
successors c42

with the derived type of V1 being (VALUES (MOD 17592186040321)
&OPTIONAL) and of V36 being (VALUES INTEGER &OPTIONAL). So cutting
cannot be performed. Replacing AMOUNT1 and AMOUNT2 with their values
and constant folding gives no new type information for V1, so Python
sees no need for reoptimization of '39>'. The problem is that Python
reoptimization mechanism is local: it does not recognize patterns,
touching several nodes.

(Someone doing next commit to the main branch is welcome to put the
test case to OPTIMIZATIONS.)

-- 
Regards,
Alexey Dejneka

"Alas, the spheres of truth are less transparent than those of
illusion." -- L.E.J. Brouwer