| blob | 45b5463330ee21f30352616c48d4d4eaf6c1ded0 |
1 #1
2 (defun mysl (s)
3 (declare (simple-string s))
4 (declare (optimize (speed 3) (safety 0) (debug 0)))
5 (let ((c 0))
6 (declare (fixnum c))
7 (dotimes (i (length s))
8 (when (eql (aref s i) #\1)
9 (incf c)))
10 c))
12 * On X86 I is represented as a tagged integer.
14 * Unnecessary move:
15 3: SLOT S!11[EDX] {SB-C::VECTOR-LENGTH 1 7} => t23[EAX]
16 4: MOVE t23[EAX] => t24[EBX]
18 --------------------------------------------------------------------------------
19 #2
20 (defun quux (v)
21 (declare (optimize (speed 3) (safety 0) (space 2) (debug 0)))
22 (declare (type (simple-array double-float 1) v))
23 (let ((s 0d0))
24 (declare (type double-float s))
25 (dotimes (i (length v))
26 (setq s (+ s (aref v i))))
27 s))
29 * Python does not combine + with AREF, so generates extra move and
30 allocates a register.
32 * On X86 Python thinks that all FP registers are directly accessible
33 and emits costy MOVE ... => FR1.
35 --------------------------------------------------------------------------------
36 #3
37 (defun bar (n)
38 (declare (optimize (speed 3) (safety 0) (space 2))
39 (type fixnum n))
40 (let ((v (make-list n)))
41 (setq v (make-array n))
42 (length v)))
44 * IR1 does not optimize away (MAKE-LIST N).
45 --------------------------------------------------------------------------------
46 #4
47 (defun bar (v1 v2)
48 (declare (optimize (speed 3) (safety 0) (space 2))
49 (type (simple-array base-char 1) v1 v2))
50 (dotimes (i (length v1))
51 (setf (aref v2 i) (aref v1 i))))
53 VOP DATA-VECTOR-SET/SIMPLE-STRING V2!14[EDI] t32[EAX] t30[S2]>t33[CL]
54 => t34[S2]<t35[AL]
55 MOV #<TN t33[CL]>, #<TN t30[S2]>
56 MOV BYTE PTR [EDI+EAX+1], #<TN t33[CL]>
57 MOV #<TN t35[AL]>, #<TN t33[CL]>
58 MOV #<TN t34[S2]>, #<TN t35[AL]>
60 * The value of DATA-VECTOR-SET is not used, so there is no need in the
61 last two moves.
63 * And why two moves?
64 --------------------------------------------------------------------------------
65 #8
66 (defun foo (d)
67 (declare (optimize (speed 3) (safety 0) (debug 0)))
68 (declare (type (double-float 0d0 1d0) d))
69 (loop for i fixnum from 1 to 5
70 for x1 double-float = (sin d) ;;; !!!
71 do (loop for j fixnum from 1 to 4
72 sum x1 double-float)))
74 Without the marked declaration Python will use boxed representation for X1.
76 This is equivalent to
78 (let ((x nil))
79 (setq x 0d0)
80 ;; use of X as DOUBLE-FLOAT
81 )
83 The initial binding is effectless, and without it X is of type
84 DOUBLE-FLOAT. Unhopefully, IR1 does not optimize away effectless
85 SETs/bindings, and IR2 does not perform type inference.
86 --------------------------------------------------------------------------------
87 #9 "Multi-path constant folding"
88 (defun foo (x)
89 (if (= (cond ((irgh x) 0)
90 ((buh x) 1)
91 (t 2))
92 0)
93 :yes
94 :no))
96 This code could be optimized to
98 (defun foo (x)
99 (cond ((irgh x) :yes)
100 ((buh x) :no)
101 (t :no)))
102 --------------------------------------------------------------------------------
103 #11
104 (inverted variant of #9)
106 (lambda (x)
107 (let ((y (sap-alien x c-string)))
108 (list (alien-sap y)
109 (alien-sap y))))
111 It could be optimized to
113 (lambda (x) (list x x))
115 (if Y were used only once, the current compiler would optimize it)
116 --------------------------------------------------------------------------------
117 #12
118 (typep (truly-the (simple-array * (*)) x) 'simple-vector)
120 tests lowtag.
121 --------------------------------------------------------------------------------
122 #13
123 FAST-+/FIXNUM and similar should accept unboxed arguments in interests
124 of representation selection. Problem: inter-TN dependencies.
125 --------------------------------------------------------------------------------
126 #14
127 The derived type of (/ (THE (DOUBLE-FLOAT (0D0)) X) (THE (DOUBLE-FLOAT
128 1D0) Y)) is (DOUBLE-FLOAT 0.0d0). While it might be reasonable, it is
129 better to derive (OR (MEMBER 0.0d0) (DOUBLE-FLOAT (0.0d0))).
130 --------------------------------------------------------------------------------
131 #15
132 On the alpha, the system is reluctant to refer directly to a constant bignum,
133 preferring to load a large constant through a slow sequence of instructions,
134 then cons up a bignum for it:
136 (LAMBDA (A)
137 (DECLARE (OPTIMIZE (SAFETY 1) (SPEED 3) (DEBUG 1))
138 (TYPE (INTEGER -10000 10000) A)
139 (IGNORABLE A))
140 (CASE A
141 ((89 125 16) (ASH A (MIN 18 -706)))
142 (T (DPB -3 (BYTE 30 30) -1))))
143 --------------------------------------------------------------------------------
144 #16
145 (do ((i 0 (1+ i)))
146 ((= i (the (integer 0 100) n)))
147 ...)
149 It is commonly expected for Python to derive (FIXNUMP I). (If ``='' is
150 replaced with ``>='', Python will do.)
151 --------------------------------------------------------------------------------
152 #17
153 Type tests for (ARRAY BIT), (ARRAY T) and similar go through full
154 %TYPEP, even though it is relatively simple to establish the arrayness
155 of an object and also to obtain the element type of an array. As of
156 sbcl-0.8.12.30, this affects at least DUMP-OBJECT through
157 COMPOUND-OBJECT-P, and (LABELS MAYBE-EMIT-MAKE-LOAD-FORMS GROVEL)
158 through TYPEP UNBOXED-ARRAY, within the compiler itself.
159 --------------------------------------------------------------------------------
160 #18
161 (lambda (x) (declare (null x)) (sxhash x)) goes through SYMBOL-HASH
162 rather than either constant-folding or manipulating NIL-VALUE or
163 NULL-TN directly.
164 --------------------------------------------------------------------------------
165 #19
166 (let ((dx (if (foo)
167 (list x)
168 (list y z))))
169 (declare (dynamic-extent dx))
170 ...)
172 DX is not allocated on stack.
173 --------------------------------------------------------------------------------
174 #20
175 (defun-with-dx foo (x)
176 (flet ((make (x)
177 (let ((l (list nil nil)))
178 (setf (first l) x)
179 (setf (second l) (1- x))
180 l)))
181 (let ((l (make x)))
182 (declare (dynamic-extent l))
183 (mapc #'print l))))
185 Result of MAKE is not stack allocated, which means that
186 stack-allocation of structures is impossible.
187 --------------------------------------------------------------------------------
188 #21
189 (defun-with-dx foo ()
190 (let ((dx (list (list 1 2) (list 3 4))))
191 (declare (dynamic-extent dx))
192 ...))
194 External list in DX is allocated on stack, but internal are not.
195 --------------------------------------------------------------------------------
196 #22
197 IR2 does not perform unused code flushing.
198 --------------------------------------------------------------------------------
199 #23
200 Python does not know that &REST lists are LISTs (and cannot derive it).
201 --------------------------------------------------------------------------------
202 #24
203 a. Iterations on &REST lists, returning them as VALUES could be
204 rewritten with &MORE vectors.
205 b. Implement local unknown-values mv-call (useful for fast type checking).
206 --------------------------------------------------------------------------------
207 #26
208 SBCL cannot derive upper bound for I and uses generic arithmetic here:
210 (defun foo (l)
211 (declare (vector l))
212 (dotimes (i (length l))
213 (if (block nil
214 (map-foo (lambda (x) (if x (return t)))
215 l))
216 t
217 nil)))
219 (So the constraint propagator or a possible future SSA-convertor
220 should know the connection between an NLE and its CLEANUP.)
221 --------------------------------------------------------------------------------
222 #27
223 Initialization of stack-allocated arrays is inefficient: we always
224 fill the vector with zeroes, even when it is not needed (as for
225 platforms with conservative GC or for arrays of unboxed objectes) and
226 is performed later explicitely.
228 (This is harder than it might look at first glance, as MAKE-ARRAY is smart
229 enough to eliminate something like ':initial-element 0'. Such an optimization
230 is valid if the vector is being allocated in the heap, but not if it is being
231 allocated on the stack. You could remove this optimization, but that makes
232 the heap-allocated case somewhat slower...)
233 --------------------------------------------------------------------------------
234 #28
235 a. Accessing raw slots in structure instances is more inefficient than
236 it could be; if we placed raw slots before the header word, we would
237 not need to do arithmetic at runtime to access them. (But beware:
238 this would complicate handling of the interior pointer).
240 b. (Also note that raw slots are currently disabled on HPPA)
241 --------------------------------------------------------------------------------
242 #29
243 Python is overly zealous when converting high-level CL functions, such
244 as MIN/MAX, LOGBITP, and LOGTEST, to low-level CL functions. Reducing
245 Python's aggressiveness would make it easier to effect changes such as
247 x86-64:
248 * direct MIN/MAX on {SINGLE,DOUBLE}-FLOATs ({MIN,MAX}S{S,D})
250 x86-64:
251 * direct LOGBITP on word-sized integers and fixnums (BT + JC)
253 x86{,-64}/PPC:
254 * branch-free MIN/MAX on word-sized integers and fixnums (floats could
255 be handled too, modulo safety considerations on the PPC)
257 x86-64:
258 * efficient LOGTESTs on word-sized integers and fixnums (TEST)
260 etc., etc.
262 (The framework for this has been implemented as of 0.9.9.18; see the
263 vm-support-routine COMBINATION-IMPLEMENTATION-STYLE and its use in
264 src/compiler/ir1opt.lisp, IR1-OPTIMIZE-COMBINATION. The above
265 optimizations are left as an exercise for the reader.)
266 --------------------------------------------------------------------------------
267 #30
268 (defun foo (x y)
269 (< x y))
271 FOO's IR1 representation is roughly:
273 (defun foo (x y)
274 (if (< x y)
275 T
276 NIL))
278 However, if a full call is generated for < (and similarly for other
279 predicate functions), then the IF is unnecessary, since the return value
280 of (< x y) is already T or NIL.
281 --------------------------------------------------------------------------------
282 #31
283 The typecheck generated for a declaration like (integer 0 45) on x86 looks
284 like:
286 ; 12B: F6C203 TEST DL, 3
287 ; 12E: 753B JNE L1
288 ; 130: 8BC2 MOV EAX, EDX
289 ; 132: 83F800 CMP EAX, 0
290 ; 135: 7C34 JL L1
291 ; 137: 8BC2 MOV EAX, EDX
292 ; 139: 3DB4000000 CMP EAX, 180
293 ; 13E: 7F2B JNLE L1
295 A better code sequence for this would be:
297 TEST DL, 3
298 JNE L1
299 MOV EAX, EDX
300 CMP EAX, 180
301 JBE L1
303 Doing an unsigned comparison means that, similarly to %CHECK-BOUND, we can
304 combine the <0 and >=bound tests. This sort of test is generated often
305 in SBCL and any array-based code that's serious about type-checking its
306 indices.
307 --------------------------------------------------------------------------------
308 #32
309 The code for a vector bounds check on x86 (similarly on x86-64) where
310 the vector is in EDX and the index in EAX looks like:
312 ; 49: L0: 8B5AFD MOV EBX, [EDX-3]
313 ; 4C: 39C3 CMP EBX, EAX
314 ; 4E: 7632 JBE L2
316 because %CHECK-BOUND is used for bounds-checking any array dimension.
317 A more efficient specialization (%CHECK-BOUND/VECTOR) would produce:
319 CMP [EDX-3], EAX
320 JBE L2
322 Which is slightly shorter and avoids using a register.
323 --------------------------------------------------------------------------------
324 #33
325 Reports from the Java camp indicate that using an SSE2-based
326 floating-point backend on x86 when possible is highly preferable to
327 using the x86 FP stack. It would be nice if SBCL included an SSE2-based
328 floating point backend with a compile-time option to switch between the
329 two.
330 --------------------------------------------------------------------------------
331 #34
332 Compiling
334 (defun foo (x y)
335 (declare (type (integer 0 45) x y))
336 (+ x y))
338 results in the following error trapping code for type-checking the
339 arguments:
341 ; 424: L0: 8B058CE31812 MOV EAX, [#x1218E38C] ; '(MOD 46)
342 ; 42A: 0F0B0A BREAK 10 ; error trap
343 ; 42D: 05 BYTE #X05
344 ; 42E: 1F BYTE #X1F ; OBJECT-NOT-TYPE-ERROR
345 ; 42F: FECE01 BYTE #XFE, #XCE, #X01 ; EDI
346 ; 432: 0E BYTE #X0E ; EAX
347 ; 433: L1: 8B0590E31812 MOV EAX, [#x1218E390] ; '(MOD 46)
348 ; 439: 0F0B0A BREAK 10 ; error trap
349 ; 43C: 03 BYTE #X03
350 ; 43D: 1F BYTE #X1F ; OBJECT-NOT-TYPE-ERROR
351 ; 43E: 8E BYTE #X8E ; EDX
352 ; 43F: 0E BYTE #X0E ; EAX
354 Notice that '(MOD 46) has two entries in the constant vector. Having
355 one would be preferable.
356 --------------------------------------------------------------------------------
357 #35
358 Compiling
360 (defun foo (a i)
361 (declare (type simple-vector a))
362 (aref a i))
364 results in the following x86 code:
366 ; 115886E9: F7C703000000 TEST EDI, 3 ; no-arg-parsing entry point
367 ; 6EF: 7510 JNE L0
368 ; 6F1: 8BC7 MOV EAX, EDI
369 ; 6F3: 83F800 CMP EAX, 0
370 ; 6F6: 7C09 JL L0
371 ; 6F8: 8BC7 MOV EAX, EDI
372 ; 6FA: 3DF8FFFF7F CMP EAX, 2147483640
373 ; 6FF: 7E0F JLE L1
374 ; 701: L0: 8B057C865811 MOV EAX, [#x1158867C] ; '(MOD
375 ; 536870911)
376 ; 707: 0F0B0A BREAK 10 ; error trap
377 ; 70A: 05 BYTE #X05
378 ; 70B: 1F BYTE #X1F ; OBJECT-NOT-TYPE-ERROR
379 ; 70C: FECE01 BYTE #XFE, #XCE, #X01 ; EDI
380 ; 70F: 0E BYTE #X0E ; EAX
381 ; 710: L1: 8B42FD MOV EAX, [EDX-3]
382 ; 713: 8BCF MOV ECX, EDI
383 ; 715: 39C8 CMP EAX, ECX
384 ; 717: 7620 JBE L2
385 ; 719: 8B540A01 MOV EDX, [EDX+ECX+1]
387 ... plus the standard return sequence and some error blocks. The
388 `TEST EDI, 3' and associated comparisons are to ensure that `I' is a
389 positive fixnum. The associated comparisons are unnecessary, as the
390 %CHECK-BOUND VOP only requires its tested index to be a fixnum and takes
391 care of the negative fixnum case itself.
393 {HAIRY-,}DATA-VECTOR-REF are DEFKNOWN'd with EXPLICIT-CHECK, which would
394 seem to take care of this, but EXPLICIT-CHECK only seems to be used when
395 compiling calls to unknown functions or similar. Furthermore,
396 EXPLICIT-CHECK, as NJF understands it, doesn't have the right
397 semantics--it suppresses all type checking of arguments, whereas what we
398 really want is to ensure that the argument is a fixnum, but not check
399 its positiveness.
400 --------------------------------------------------------------------------------
401 #36
403 In #35, the CMP EAX, $foo instructions are all preceded by a MOV. They
404 appear to be unnecessary, but are necessary because in IR2, EDI is a
405 DESCRIPTOR-REG, whereas EAX is an ANY-REG--and the comparison VOPs only
406 accept ANY-REGs. Therefore, the MOVs are "necessary" to ensure that the
407 comparison VOP receives an TN of the appropriate storage class.
409 Obviously, it would be better if a) we only performed one MOV prior to
410 all three comparisons or b) eliminated the necessity of the MOV(s)
411 altogether. The former option is probably easier than the latter.
413 --------------------------------------------------------------------------------
414 #37
416 Dynamic extent allocation doesn't currently work for one-element lists,
417 since there's a source transform from (LIST X) to (CONS X NIL).
419 --------------------------------------------------------------------------------
420 #38
422 (setf (subseq s1 start1 end1) (subseq s2 start2 end1))
424 could be transformed into
426 (let ((#:s2 s2)
427 (#:start2 start2)
428 (#:end2 end2))
429 (replace s1 #:s2 :start1 start1 :end1 end1 :start2 #:start2 :end2 #:end2))
431 when the return value is unused, avoiding the need to cons up the new sequence.
433 --------------------------------------------------------------------------------
434 #39
436 (let ((*foo* 42)) ...)
438 currently compiles to code that ensures the TLS index at runtime, which
439 is both a decently large chunk of code and unnecessary, as we could ensure
440 the TLS index at load-time as well.