OPTIMIZATIONS

#1
(defun mysl (s)
    (declare (simple-string s))
    (declare (optimize (speed 3) (safety 0) (debug 0)))
    (let ((c 0))
      (declare (fixnum c))
      (dotimes (i (length s))
        (when (eql (aref s i) #\1)
          (incf c)))
      c))

* On X86 I is represented as a tagged integer.

* Unnecessary move:
  3: SLOT S!11[EDX] {SB-C::VECTOR-LENGTH 1 7} => t23[EAX]
  4: MOVE t23[EAX] => t24[EBX]

--------------------------------------------------------------------------------
#2
(defun quux (v)
  (declare (optimize (speed 3) (safety 0) (space 2) (debug 0)))
  (declare (type (simple-array double-float 1) v))
  (let ((s 0d0))
    (declare (type double-float s))
    (dotimes (i (length v))
      (setq s (+ s (aref v i))))
    s))

* Python does not combine + with AREF, so generates extra move and
  allocates a register.

* On X86 Python thinks that all FP registers are directly accessible
  and emits costy MOVE ... => FR1.

--------------------------------------------------------------------------------
#3
(defun bar (n)
  (declare (optimize (speed 3) (safety 0) (space 2))
           (type fixnum n))
  (let ((v (make-list n)))
    (setq v (make-array n))
    (length v)))

* IR1 does not optimize away (MAKE-LIST N).
--------------------------------------------------------------------------------
#4
(defun bar (v1 v2)
  (declare (optimize (speed 3) (safety 0) (space 2))
           (type (simple-array base-char 1) v1 v2))
  (dotimes (i (length v1))
    (setf (aref v2 i) (aref v1 i))))

VOP DATA-VECTOR-SET/SIMPLE-STRING V2!14[EDI] t32[EAX] t30[S2]>t33[CL]
                                  => t34[S2]<t35[AL] 
        MOV     #<TN t33[CL]>, #<TN t30[S2]>
        MOV     BYTE PTR [EDI+EAX+1], #<TN t33[CL]>
        MOV     #<TN t35[AL]>, #<TN t33[CL]>
        MOV     #<TN t34[S2]>, #<TN t35[AL]>

* The value of DATA-VECTOR-SET is not used, so there is no need in the
  last two moves.

* And why two moves?
--------------------------------------------------------------------------------
#8
(defun foo (d)
  (declare (optimize (speed 3) (safety 0) (debug 0)))
  (declare (type (double-float 0d0 1d0) d))
  (loop for i fixnum from 1 to 5
        for x1 double-float = (sin d) ;;; !!!
        do (loop for j fixnum from 1 to 4
                 sum x1 double-float)))

Without the marked declaration Python will use boxed representation for X1.

This is equivalent to

(let ((x nil))
  (setq x 0d0)
  ;; use of X as DOUBLE-FLOAT
)

The initial binding is effectless, and without it X is of type
DOUBLE-FLOAT. Unhopefully, IR1 does not optimize away effectless
SETs/bindings, and IR2 does not perform type inference.
--------------------------------------------------------------------------------
#9 "Multi-path constant folding"
(defun foo (x)
  (if (= (cond ((irgh x) 0)
               ((buh x) 1)
               (t 2))
         0)
      :yes
      :no))

This code could be optimized to

(defun foo (x)
  (cond ((irgh x) :yes)
        ((buh x) :no)
        (t :no)))
--------------------------------------------------------------------------------
#11
(inverted variant of #9)

(lambda (x)
  (let ((y (sap-alien x c-string)))
    (list (alien-sap y)
          (alien-sap y))))

It could be optimized to

(lambda (x) (list x x))

(if Y were used only once, the current compiler would optimize it)
--------------------------------------------------------------------------------
#12
(typep (truly-the (simple-array * (*)) x) 'simple-vector)

tests lowtag.
--------------------------------------------------------------------------------
#13
FAST-+/FIXNUM and similar should accept unboxed arguments in interests
of representation selection. Problem: inter-TN dependencies.
--------------------------------------------------------------------------------
#14
The derived type of (/ (THE (DOUBLE-FLOAT (0D0)) X) (THE (DOUBLE-FLOAT
1D0) Y)) is (DOUBLE-FLOAT 0.0d0). While it might be reasonable, it is
better to derive (OR (MEMBER 0.0d0) (DOUBLE-FLOAT (0.0d0))).
--------------------------------------------------------------------------------
#15
On the alpha, the system is reluctant to refer directly to a constant bignum,
preferring to load a large constant through a slow sequence of instructions,
then cons up a bignum for it:

(LAMBDA (A)
  (DECLARE (OPTIMIZE (SAFETY 1) (SPEED 3) (DEBUG 1))
           (TYPE (INTEGER -10000 10000) A)
           (IGNORABLE A))
  (CASE A
    ((89 125 16) (ASH A (MIN 18 -706)))
    (T (DPB -3 (BYTE 30 30) -1))))
--------------------------------------------------------------------------------
#16
(do ((i 0 (1+ i)))
    ((= i (the (integer 0 100) n)))
  ...)

It is commonly expected for Python to derive (FIXNUMP I). (If ``='' is
replaced with ``>='', Python will do.)
--------------------------------------------------------------------------------
#17 
Type tests for (ARRAY BIT), (ARRAY T) and similar go through full
%TYPEP, even though it is relatively simple to establish the arrayness
of an object and also to obtain the element type of an array.  As of
sbcl-0.8.12.30, this affects at least DUMP-OBJECT through
COMPOUND-OBJECT-P, and (LABELS MAYBE-EMIT-MAKE-LOAD-FORMS GROVEL)
through TYPEP UNBOXED-ARRAY, within the compiler itself.
--------------------------------------------------------------------------------
#20
(defun-with-dx foo (x)
  (flet ((make (x)
           (let ((l (list nil nil)))
             (setf (first l) x)
             (setf (second l) (1- x))
             l)))
    (let ((l (make x)))
      (declare (dynamic-extent l))
      (mapc #'print l))))

Result of MAKE is not stack allocated.
--------------------------------------------------------------------------------
#22
IR2 does not perform unused code flushing.
--------------------------------------------------------------------------------
#24
a. Iterations on &REST lists could be rewritten with &MORE vectors.
b. Implement local unknown-values mv-call (useful for fast type checking).
--------------------------------------------------------------------------------
#26
SBCL cannot derive upper bound for I and uses generic arithmetic here:

(defun foo (l)
  (declare (vector l))
  (dotimes (i (length l))
    (if (block nil
          (map-foo (lambda (x) (if x (return t)))
                   l))
        t
        nil)))

(So the constraint propagator or a possible future SSA-convertor
should know the connection between an NLE and its CLEANUP.)
--------------------------------------------------------------------------------
#27
Initialization of stack-allocated arrays is inefficient: we always
fill the vector with zeroes, even when it is not needed (as for
platforms with conservative GC or for arrays of unboxed objectes) and
is performed later explicitely.

(This is harder than it might look at first glance, as MAKE-ARRAY is smart
enough to eliminate something like ':initial-element 0'.  Such an optimization
is valid if the vector is being allocated in the heap, but not if it is being
allocated on the stack.  You could remove this optimization, but that makes
the heap-allocated case somewhat slower...)

To do this, extend ALLOCATE-VECTOR with ALLOW-JUNK argument, and when
stack allocating don't zero if it is true -- and probably ALLOW-JUNK iff
the vector is a specialized one (cannot have pointers.)
--------------------------------------------------------------------------------
#29
Python is overly zealous when converting high-level CL functions, such
as MIN/MAX, LOGBITP, and LOGTEST, to low-level CL functions.  Reducing
Python's aggressiveness would make it easier to effect changes such as

x86-64:
* direct MIN/MAX on {SINGLE,DOUBLE}-FLOATs ({MIN,MAX}S{S,D})

x86-64:
* direct LOGBITP on word-sized integers and fixnums (BT + JC)

x86{,-64}/PPC:
* branch-free MIN/MAX on word-sized integers and fixnums (floats could
  be handled too, modulo safety considerations on the PPC)

x86-64:
* efficient LOGTESTs on word-sized integers and fixnums (TEST)

etc., etc.

(The framework for this has been implemented as of 0.9.9.18; see the
vm-support-routine COMBINATION-IMPLEMENTATION-STYLE and its use in
src/compiler/ir1opt.lisp, IR1-OPTIMIZE-COMBINATION.  The above
optimizations are left as an exercise for the reader.)
--------------------------------------------------------------------------------
#31
The typecheck generated for a declaration like (integer 0 45) on x86 looks
like:

;      12B:       F6C203           TEST DL, 3
;      12E:       753B             JNE L1
;      130:       8BC2             MOV EAX, EDX
;      132:       83F800           CMP EAX, 0
;      135:       7C34             JL L1
;      137:       8BC2             MOV EAX, EDX
;      139:       3DB4000000       CMP EAX, 180
;      13E:       7F2B             JNLE L1

A better code sequence for this would be:

  TEST DL, 3
  JNE L1
  MOV EAX, EDX
  CMP EAX, 180
  JBE L1

Doing an unsigned comparison means that, similarly to %CHECK-BOUND, we can
combine the <0 and >=bound tests.  This sort of test is generated often
in SBCL and any array-based code that's serious about type-checking its
indices.
--------------------------------------------------------------------------------
#32
The code for a vector bounds check on x86 (similarly on x86-64) where
the vector is in EDX and the index in EAX looks like:

;       49: L0:   8B5AFD           MOV EBX, [EDX-3]
;       4C:       39C3             CMP EBX, EAX
;       4E:       7632             JBE L2

because %CHECK-BOUND is used for bounds-checking any array dimension.
A more efficient specialization (%CHECK-BOUND/VECTOR) would produce:

  CMP [EDX-3], EAX
  JBE L2

Which is slightly shorter and avoids using a register.
--------------------------------------------------------------------------------
#33
Reports from the Java camp indicate that using an SSE2-based
floating-point backend on x86 when possible is highly preferable to
using the x86 FP stack.  It would be nice if SBCL included an SSE2-based
floating point backend with a compile-time option to switch between the
two.
--------------------------------------------------------------------------------
#35
Compiling

(defun foo (a i)
  (declare (type simple-vector a))
  (aref a i))

results in the following x86 code:

; 115886E9:       F7C703000000     TEST EDI, 3                ; no-arg-parsing entry point
;      6EF:       7510             JNE L0
;      6F1:       8BC7             MOV EAX, EDI
;      6F3:       83F800           CMP EAX, 0
;      6F6:       7C09             JL L0
;      6F8:       8BC7             MOV EAX, EDI
;      6FA:       3DF8FFFF7F       CMP EAX, 2147483640
;      6FF:       7E0F             JLE L1
;      701: L0:   8B057C865811     MOV EAX, [#x1158867C]      ; '(MOD
                                                              ;   536870911)
;      707:       0F0B0A           BREAK 10                   ; error trap
;      70A:       05               BYTE #X05
;      70B:       1F               BYTE #X1F                  ; OBJECT-NOT-TYPE-ERROR
;      70C:       FECE01           BYTE #XFE, #XCE, #X01      ; EDI
;      70F:       0E               BYTE #X0E                  ; EAX
;      710: L1:   8B42FD           MOV EAX, [EDX-3]
;      713:       8BCF             MOV ECX, EDI
;      715:       39C8             CMP EAX, ECX
;      717:       7620             JBE L2
;      719:       8B540A01         MOV EDX, [EDX+ECX+1]

... plus the standard return sequence and some error blocks.  The
`TEST EDI, 3' and associated comparisons are to ensure that `I' is a
positive fixnum.  The associated comparisons are unnecessary, as the
%CHECK-BOUND VOP only requires its tested index to be a fixnum and takes
care of the negative fixnum case itself.

{HAIRY-,}DATA-VECTOR-REF are DEFKNOWN'd with EXPLICIT-CHECK, which would
seem to take care of this, but EXPLICIT-CHECK only seems to be used when
compiling calls to unknown functions or similar.  Furthermore,
EXPLICIT-CHECK, as NJF understands it, doesn't have the right
semantics--it suppresses all type checking of arguments, whereas what we
really want is to ensure that the argument is a fixnum, but not check
its positiveness.
--------------------------------------------------------------------------------
#36

In #35, the CMP EAX, $foo instructions are all preceded by a MOV.  They
appear to be unnecessary, but are necessary because in IR2, EDI is a
DESCRIPTOR-REG, whereas EAX is an ANY-REG--and the comparison VOPs only
accept ANY-REGs.  Therefore, the MOVs are "necessary" to ensure that the
comparison VOP receives an TN of the appropriate storage class.

Obviously, it would be better if a) we only performed one MOV prior to
all three comparisons or b) eliminated the necessity of the MOV(s)
altogether.  The former option is probably easier than the latter.

--------------------------------------------------------------------------------
#38

(setf (subseq s1 start1 end1) (subseq s2 start2 end1))

could be transformed into

(let ((#:s2 s2)
      (#:start2 start2)
      (#:end2 end2))
 (replace s1 #:s2 :start1 start1 :end1 end1 :start2 #:start2 :end2 #:end2))

when the return value is unused, avoiding the need to cons up the new sequence.

--------------------------------------------------------------------------------
#39

(let ((*foo* 42)) ...)

currently compiles to code that ensures the TLS index at runtime, which
is both a decently large chunk of code and unnecessary, as we could ensure
the TLS index at load-time as well.
[Note that x86-64 already does this.]

--------------------------------------------------------------------------------
#40

When FTYPE is declared -- to say (function (t t t t t) t), and
function has a compiler-macro,

  (apply #'foo 'x1 x2 'x3 more)

can be transformed into

  (apply (lambda (x2 x4 x5) (foo 'x1 x2 'x3 x4 x5)) x2 more)

which allows compiler-macro-expansion for FOO. (Only constant
arguments can be moved inside the new lambda -- otherwise evaluation
order is altered.)

--------------------------------------------------------------------------------
#41

The unibyte external formats are written in a very generic way.  Three
optimizations immediately applicable that could be automatically
generated:

(a) if the external format merely permutes the first 256 characters, a
    constant-time lookup (rather than a binary search) could be
    performed on output.  This applies at least to EBCDIC, which
    currently has a hand-rolled mapper instead.

(b) if there are no undefined characters corresponding to the 256
    codes, then no error checking need be done on input.

(c) if there is a way to use particular bits of the exceptional
    characters, constant-time output (rather than binary search) can
    still be achieved as used to be done by the latin-9 external
    format before 1.0.31.