Re: [Sbcl-devel] generic mask-signed-field conses a ton despite fixnum input/output

[Paul K. <pv...@pv...>]

Paul Khuong wrote:
> Douglas Katzman wrote:
>> Another thing. Despite forcing there to be no type-check on the 'index',
>> we emit (IF blah NIL NIL),
>> respectively (IF blah T T) by substituting -1 for 0.
>>
>> * (disassemble '(lambda (x) (logbitp (truly-the unsigned-byte x) 0)))
[...]
> That's because the IF node branches to two equivalent but different
> basic blocks. It should be possible to detect simple cases (e.g., nodes
> that only deliver the same REF to the same LVAR). The logic would be
> necessary to try and be more clever when choosing templates for
> predicates (VOPs that directly jump to a target vs only computing flags).

I now see it's not immediately apparent that the branch comes from the 
transform for LOGBITP:
(deftransform logbitp
     ((index integer) (unsigned-byte (or (signed-byte #.sb!vm:n-word-bits)
                                         (unsigned-byte 
#.sb!vm:n-word-bits))))
   `(if (>= index #.sb!vm:n-word-bits)
        (minusp integer)
        (not (zerop (logand integer (ash 1 index))))))

It's a transform because inlining only happens when the second argument 
is a machine word (also, there's a slight [common] bug in the type 
specification for that argument).

Now that I think about it though, these two cases should really be 
handled in the transform. LOGBITP of 0 and -1 are trivial cases of the 
kind of type-directed simplifications we do for bitwise ops and modular 
arithmetic. A general transform on LOGBITP seems more elegant to me.

Paul Khuong