| 1 |
/* |
| 2 |
* dispatch.c - method dispatch accelerator |
| 3 |
* |
| 4 |
* Copyright (c) 2017-2024 Shiro Kawai <[email protected]> |
| 5 |
* |
| 6 |
* Redistribution and use in source and binary forms, with or without |
| 7 |
* modification, are permitted provided that the following conditions |
| 8 |
* are met: |
| 9 |
* |
| 10 |
* 1. Redistributions of source code must retain the above copyright |
| 11 |
* notice, this list of conditions and the following disclaimer. |
| 12 |
* |
| 13 |
* 2. Redistributions in binary form must reproduce the above copyright |
| 14 |
* notice, this list of conditions and the following disclaimer in the |
| 15 |
* documentation and/or other materials provided with the distribution. |
| 16 |
* |
| 17 |
* 3. Neither the name of the authors nor the names of its contributors |
| 18 |
* may be used to endorse or promote products derived from this |
| 19 |
* software without specific prior written permission. |
| 20 |
* |
| 21 |
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
| 22 |
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
| 23 |
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR |
| 24 |
* A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT |
| 25 |
* OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, |
| 26 |
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED |
| 27 |
* TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR |
| 28 |
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF |
| 29 |
* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING |
| 30 |
* NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS |
| 31 |
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 32 |
*/ |
| 33 |
|
| 34 |
#define LIBGAUCHE_BODY |
| 35 |
#include "gauche.h" |
| 36 |
#include "gauche/priv/configP.h" |
| 37 |
#include "gauche/priv/atomicP.h" |
| 38 |
#include "gauche/priv/dispatchP.h" |
| 39 |
|
| 40 |
/* |
| 41 |
* Method dispatch acceleration |
| 42 |
* |
| 43 |
* gf->dispatcher may contain a structure that accelerates method dispatch. |
| 44 |
* The structure must be treated as opaque from other parts. |
| 45 |
* |
| 46 |
* In the current implementation, we use special open-addressing hash table. |
| 47 |
* We don't use ScmHashCore for a few reasons: |
| 48 |
* |
| 49 |
* - We mutex modification of the table by gf->mutex, but we don't want |
| 50 |
* to lock for searching it. ScmHashCore doesn't guarantee consistency |
| 51 |
* of reading while modifying. |
| 52 |
* - It is in performance critical path, and we can take advantage of |
| 53 |
* domain knowledge to make it faster than generic implementation. |
| 54 |
* |
| 55 |
* At this moment, the dispatch accelerator isn't built automatically. |
| 56 |
* You need to call gauche.object#generic-build-dispatcher! explicitly |
| 57 |
* on a generic function. The effect varies depending on the methods |
| 58 |
* the GF has, so we selectively turn on this feature for speed-sensitive |
| 59 |
* GFs and see how it goes. If it is proven to be effective enough, |
| 60 |
* we might implement an automated mechanism when GF meets certain |
| 61 |
* criteria. |
| 62 |
* |
| 63 |
* We take advantage of the following facts: |
| 64 |
* |
| 65 |
* - Once we hit the leaf method we don't need to find less-specific |
| 66 |
* methods at all. |
| 67 |
* - Some generic functions has methods most of which are leaf methods, |
| 68 |
* and can be mostly specialized by a single argument. The typical |
| 69 |
* example is `ref'. |
| 70 |
* |
| 71 |
* When methods are mostly specilized by N-th argument, we call the N |
| 72 |
* as AXIS (zero based). Typically, methods tend to be specialized by |
| 73 |
* the first argument, in which case axis is 0. |
| 74 |
* |
| 75 |
* The method dispatcher contains a hash table, where each key is |
| 76 |
* a tuple (<class>, <number-of-args>), and its value is a list of |
| 77 |
* methods, whose axis specializer is <class> and which can take |
| 78 |
* <number-of-args>. |
| 79 |
* |
| 80 |
* When a GF with dispatch accelerator is called, we retrieve the class |
| 81 |
* of the axis of the actual argument, and lookup the hash table with |
| 82 |
* that class and the number of actual arguments. If we have a hit, |
| 83 |
* and the entry solely consists of leaf methods, |
| 84 |
* the list of methods are the list of applicable methods---since there |
| 85 |
* are no more specific methods, and we don't need to look for less specific |
| 86 |
* methods. This means... |
| 87 |
* |
| 88 |
* - We don't need to go through the entire method list to find |
| 89 |
* all applicable methods. |
| 90 |
* - We don't need to allocate to construct the list of applicable |
| 91 |
* methods. |
| 92 |
* |
| 93 |
* Note that if the entry has at least one non-leaf method, we have to |
| 94 |
* fall back to the normal path, since there may be other applicable |
| 95 |
* methods that might be called via next-method. To cut the overhead |
| 96 |
* of checking, we keep the list of leaf methods and the one of non-leaf |
| 97 |
* methods separately. |
| 98 |
*/ |
| 99 |
|
| 100 |
/* On concurrent access: |
| 101 |
* |
| 102 |
* We don't want to lock the method hash for every invocation of GF. |
| 103 |
* So we employ some atomic pointer operations so that readers can |
| 104 |
* see the consistent state even another thread is modifying it. |
| 105 |
* |
| 106 |
* We still acquire GF lock whenever we modify the mhash, so there's |
| 107 |
* at most one thread that's modifying it. Modification is only required |
| 108 |
* when a new method is added to the GF or any one of classes that |
| 109 |
* specializes one of the methods is redefined. |
| 110 |
* |
| 111 |
* The modification is done by making a modified copy of (sub)structure |
| 112 |
* and then swap the pointer atomically, thus the concurrent reader |
| 113 |
* won't see inconsistent state. |
| 114 |
* |
| 115 |
* The locking on GF is done in class.c. |
| 116 |
*/ |
| 117 |
|
| 118 |
struct ScmMethodDispatcherRec { |
| 119 |
int axis; /* Which argument we look at? |
| 120 |
This is immutable. */ |
| 121 |
ScmAtomicVar methodHash; /* mhash. In case mhash is extended, |
| 122 |
we atomically swap reference. */ |
| 123 |
}; |
| 124 |
|
| 125 |
typedef struct mhash_entry_rec { |
| 126 |
ScmClass *klass; |
| 127 |
int nargs; |
| 128 |
ScmObj leaves; /* list of matching leaf methods */ |
| 129 |
ScmObj nonleaves; /* list of matching non-leaf methods */ |
| 130 |
} mhash_entry; |
| 131 |
|
| 132 |
typedef struct mhash_rec { |
| 133 |
int size; /* # of bins. power of 2. */ |
| 134 |
int num_entries; /* # of active entries. */ |
| 135 |
ScmAtomicVar bins[1]; /* Table. Each entry may have one of: |
| 136 |
0 - free |
| 137 |
1 - deleted |
| 138 |
mhash_entry |
| 139 |
We might atomically change the bin value, |
| 140 |
which shouldn't affect concurrently reading |
| 141 |
threads. |
| 142 |
*/ |
| 143 |
} mhash; |
| 144 |
|
| 145 |
|
| 146 |
static inline u_long mhashfn(ScmClass *k, int nargs) |
| 147 |
{ |
| 148 |
return (((SCM_WORD(k) >> 3) + nargs) * 2654435761UL) >> 20; |
| 149 |
} |
| 150 |
|
| 151 |
static mhash *make_mhash(int size) |
| 152 |
{ |
| 153 |
mhash *mh = SCM_NEW2(mhash*, sizeof(mhash)+(sizeof(ScmAtomicWord)*(size-1))); |
| 154 |
mh->size = size; |
| 155 |
mh->num_entries = 0; |
| 156 |
for (int i=0; i<size; i++) mh->bins[i] = 0; |
| 157 |
return mh; |
| 158 |
} |
| 159 |
|
| 160 |
static ScmObj mhash_probe(const mhash *h, ScmClass *k, int nargs) |
| 161 |
{ |
| 162 |
/* Quadratic probing |
| 163 |
j = H_i(k,nargs) + (i + i^2)/2 |
| 164 |
H_{i+1} - H_i = ((i+1) + (i+1)^2 - (i + i^2))/2 = i + 1 |
| 165 |
*/ |
| 166 |
u_long j = mhashfn(k, nargs) & (h->size - 1); |
| 167 |
int i = 0; |
| 168 |
for (; i < h->size; i++) { |
| 169 |
/* Need to strip 'const', because of C11 error |
| 170 |
http://www.open-std.org/jtc1/sc22/wg14/www/docs/summary.htm#dr_459 */ |
| 171 |
ScmAtomicVar *loc = (ScmAtomicVar*)&h->bins[j]; |
| 172 |
ScmWord w = SCM_WORD(Scm_AtomicLoad(loc)); |
| 173 |
if (w == 0) break; |
| 174 |
if (w != 1) { |
| 175 |
mhash_entry *e = (mhash_entry*)w; |
| 176 |
if (e->klass == k && e->nargs == nargs) { |
| 177 |
if (SCM_NULLP(e->nonleaves)) return e->leaves; |
| 178 |
else return SCM_FALSE; |
| 179 |
} |
| 180 |
} |
| 181 |
j = (j + i + 1) & (h->size - 1); |
| 182 |
} |
| 183 |
return SCM_FALSE; |
| 184 |
} |
| 185 |
|
| 186 |
static mhash *mhash_insert_1(mhash *h, ScmClass *k, int nargs, ScmMethod *m) |
| 187 |
{ |
| 188 |
u_long j = mhashfn(k, nargs) & (h->size - 1); |
| 189 |
long free_slot = -1; |
| 190 |
ScmObj ltail = SCM_NIL, ntail = SCM_NIL; |
| 191 |
int i = 0; |
| 192 |
for (; i < h->size; i++) { |
| 193 |
ScmWord w = SCM_WORD(Scm_AtomicLoad(&h->bins[j])); |
| 194 |
if (w == 0) { /* end of chain */ |
| 195 |
if (free_slot < 0) free_slot = j; |
| 196 |
break; |
| 197 |
} |
| 198 |
if (w == 1) { |
| 199 |
if (free_slot < 0) free_slot = j; |
| 200 |
continue; |
| 201 |
} |
| 202 |
mhash_entry *e = (mhash_entry*)w; |
| 203 |
if (e->klass == k && e->nargs == nargs) { |
| 204 |
free_slot = j; |
| 205 |
ltail = e->leaves; |
| 206 |
ntail = e->nonleaves; |
| 207 |
h->num_entries--; |
| 208 |
break; |
| 209 |
} |
| 210 |
j = (j + i + 1) & (h->size - 1); |
| 211 |
|
| 212 |
} |
| 213 |
SCM_ASSERT(free_slot >= 0); |
| 214 |
mhash_entry *e = SCM_NEW(mhash_entry); |
| 215 |
e->klass = k; |
| 216 |
e->nargs = nargs; |
| 217 |
e->leaves = SCM_METHOD_LEAF_P(m)? Scm_Cons(SCM_OBJ(m), ltail) : ltail; |
| 218 |
e->nonleaves = SCM_METHOD_LEAF_P(m) ? ntail : Scm_Cons(SCM_OBJ(m), ntail); |
| 219 |
Scm_AtomicStoreFull(&h->bins[free_slot], (ScmAtomicWord)e); |
| 220 |
h->num_entries++; |
| 221 |
return h; |
| 222 |
} |
| 223 |
|
| 224 |
static mhash *mhash_insert(mhash *h, ScmClass *k, int nargs, ScmMethod *m) |
| 225 |
{ |
| 226 |
if (h->size <= h->num_entries*2) { |
| 227 |
/* extend */ |
| 228 |
mhash *nh = make_mhash(h->size*2); |
| 229 |
nh->num_entries = h->num_entries; |
| 230 |
for (int i = 0; i < h->size; i++) { |
| 231 |
ScmWord w = h->bins[i]; |
| 232 |
if (w == 0 || w == 1) continue; |
| 233 |
mhash_entry *e = (mhash_entry*)w; |
| 234 |
u_long j = mhashfn(e->klass, e->nargs) & (nh->size - 1); |
| 235 |
int k = 0; |
| 236 |
for (; k < nh->size; k++) { |
| 237 |
if (SCM_WORD(nh->bins[j]) == 0) { |
| 238 |
nh->bins[j] = w; |
| 239 |
break; |
| 240 |
} |
| 241 |
j = (j + k + 1) & (nh->size - 1); |
| 242 |
} |
| 243 |
SCM_ASSERT(k < nh->size); |
| 244 |
} |
| 245 |
h = nh; |
| 246 |
} |
| 247 |
return mhash_insert_1(h, k, nargs, m); |
| 248 |
} |
| 249 |
|
| 250 |
static mhash *mhash_delete(mhash *h, ScmClass *k, int nargs, ScmMethod *m) |
| 251 |
{ |
| 252 |
u_long j = mhashfn(k, nargs) & (h->size - 1); |
| 253 |
|
| 254 |
int i = 0; |
| 255 |
for (; i < h->size; i++) { |
| 256 |
ScmWord w = SCM_WORD(Scm_AtomicLoad(&h->bins[j])); |
| 257 |
if (w == 0) break; |
| 258 |
if (w == 1) continue; |
| 259 |
mhash_entry *e = (mhash_entry*)w; |
| 260 |
if (e->klass == k && e->nargs == nargs) { |
| 261 |
ScmObj ml = e->leaves; |
| 262 |
ScmObj mn = e->nonleaves; |
| 263 |
if (SCM_PAIRP(ml) && SCM_EQ(SCM_CAR(ml), SCM_OBJ(m))) { |
| 264 |
ml = SCM_CDR(ml); /* fast path */ |
| 265 |
} else { |
| 266 |
ml = Scm_Delete(SCM_OBJ(m), ml, SCM_CMP_EQ); |
| 267 |
} |
| 268 |
if (SCM_PAIRP(mn) && SCM_EQ(SCM_CAR(mn), SCM_OBJ(m))) { |
| 269 |
mn = SCM_CDR(mn); /* fast path */ |
| 270 |
} else { |
| 271 |
mn = Scm_Delete(SCM_OBJ(m), mn, SCM_CMP_EQ); |
| 272 |
} |
| 273 |
|
| 274 |
if (SCM_NULLP(ml) && SCM_NULLP(ml)) { |
| 275 |
h->num_entries--; |
| 276 |
Scm_AtomicStore(&h->bins[j], 1); /* mark as deleted */ |
| 277 |
} else { |
| 278 |
mhash_entry *e = SCM_NEW(mhash_entry); |
| 279 |
e->klass = k; |
| 280 |
e->nargs = nargs; |
| 281 |
e->leaves = ml; |
| 282 |
e->nonleaves = mn; |
| 283 |
Scm_AtomicStoreFull(&h->bins[j], (ScmAtomicWord)e); |
| 284 |
} |
| 285 |
break; |
| 286 |
} |
| 287 |
j = (j + i + 1) & (h->size - 1); |
| 288 |
} |
| 289 |
return h; |
| 290 |
} |
| 291 |
|
| 292 |
static void mhash_print(mhash *h, ScmPort *out) |
| 293 |
{ |
| 294 |
Scm_Printf(out, "mhash size=%d num_entries=%d\n", h->size, h->num_entries); |
| 295 |
for (int i=0; i<h->size; i++) { |
| 296 |
ScmWord w = SCM_WORD(h->bins[i]); |
| 297 |
if (w == 0) { |
| 298 |
Scm_Printf(out, "[%3d] empty\n\n\n", i); |
| 299 |
} else if (w == 1) { |
| 300 |
Scm_Printf(out, "[%3d] deleted\n\n\n", i); |
| 301 |
} else { |
| 302 |
mhash_entry *e = (mhash_entry*)w; |
| 303 |
Scm_Printf(out, "[%3d] %lu %S(%d)\n", i, |
| 304 |
(mhashfn(e->klass, e->nargs) % h->size), |
| 305 |
e->klass, e->nargs); |
| 306 |
Scm_Printf(out, " Leaves: %S\n", e->leaves); |
| 307 |
Scm_Printf(out, " NonLeaves:%S\n", e->nonleaves); |
| 308 |
} |
| 309 |
} |
| 310 |
} |
| 311 |
|
| 312 |
static mhash *add_method_to_dispatcher(mhash *h, int axis, ScmMethod *m) |
| 313 |
{ |
| 314 |
int req = SCM_PROCEDURE_REQUIRED(m); |
| 315 |
if (req >= axis) { |
| 316 |
ScmClass *klass = m->specializers[axis]; |
| 317 |
if (SCM_PROCEDURE_OPTIONAL(m)) { |
| 318 |
for (int k = req; k < SCM_DISPATCHER_MAX_NARGS; k++) |
| 319 |
h = mhash_insert(h, klass, k, m); |
| 320 |
} else { |
| 321 |
h = mhash_insert(h, klass, req, m); |
| 322 |
} |
| 323 |
} |
| 324 |
return h; |
| 325 |
} |
| 326 |
|
| 327 |
static mhash *delete_method_from_dispatcher(mhash *h, int axis, ScmMethod *m) |
| 328 |
{ |
| 329 |
int req = SCM_PROCEDURE_REQUIRED(m); |
| 330 |
if (req >= axis) { |
| 331 |
ScmClass *klass = m->specializers[axis]; |
| 332 |
if (SCM_PROCEDURE_OPTIONAL(m)) { |
| 333 |
for (int k = req; k < SCM_DISPATCHER_MAX_NARGS; k++) |
| 334 |
h = mhash_delete(h, klass, k, m); |
| 335 |
} else { |
| 336 |
h = mhash_delete(h, klass, req, m); |
| 337 |
} |
| 338 |
} |
| 339 |
return h; |
| 340 |
} |
| 341 |
|
| 342 |
/* |
| 343 |
NB: We run through the method list twice, first process the |
| 344 |
leaf methods, and then process non-leaf methods. Non-leaf methods |
| 345 |
cancels the dispatcher entry and forces to go through normal route. |
| 346 |
*/ |
| 347 |
ScmMethodDispatcher *Scm__BuildMethodDispatcher(ScmObj methods, int axis) |
| 348 |
{ |
| 349 |
mhash *mh = make_mhash(32); |
| 350 |
ScmObj mm; |
| 351 |
for (int i = 0; i < 2; i++) { |
| 352 |
SCM_FOR_EACH(mm, methods) { |
| 353 |
ScmMethod *m = SCM_METHOD(SCM_CAR(mm)); |
| 354 |
if ((i == 0 && SCM_METHOD_LEAF_P(m)) |
| 355 |
|| (i == 1 && !SCM_METHOD_LEAF_P(m))) { |
| 356 |
mh = add_method_to_dispatcher(mh, axis, m); |
| 357 |
} |
| 358 |
} |
| 359 |
} |
| 360 |
ScmMethodDispatcher *dis = SCM_NEW(ScmMethodDispatcher); |
| 361 |
dis->axis = axis; |
| 362 |
dis->methodHash = (ScmAtomicWord)mh; |
| 363 |
return dis; |
| 364 |
} |
| 365 |
|
| 366 |
void Scm__MethodDispatcherAdd(ScmMethodDispatcher *dis, ScmMethod *m) |
| 367 |
{ |
| 368 |
mhash *h = (mhash*)Scm_AtomicLoad(&dis->methodHash); |
| 369 |
mhash *h2 = add_method_to_dispatcher(h, dis->axis, m); |
| 370 |
if (h != h2) Scm_AtomicStore(&dis->methodHash, (ScmAtomicWord)h2); |
| 371 |
} |
| 372 |
|
| 373 |
void Scm__MethodDispatcherDelete(ScmMethodDispatcher *dis, ScmMethod *m) |
| 374 |
{ |
| 375 |
mhash *h = (mhash*)Scm_AtomicLoad(&dis->methodHash); |
| 376 |
mhash *h2 = delete_method_from_dispatcher(h, dis->axis, m); |
| 377 |
if (h != h2) Scm_AtomicStore(&dis->methodHash, (ScmAtomicWord)h2); |
| 378 |
} |
| 379 |
|
| 380 |
ScmObj Scm__MethodDispatcherLookup(ScmMethodDispatcher *dis, |
| 381 |
ScmClass **typev, int argc) |
| 382 |
{ |
| 383 |
if (dis->axis <= argc) { |
| 384 |
ScmClass *selector = typev[dis->axis]; |
| 385 |
mhash *h = (mhash*)Scm_AtomicLoad(&dis->methodHash); |
| 386 |
return mhash_probe(h, selector, argc); |
| 387 |
} else { |
| 388 |
return SCM_FALSE; |
| 389 |
} |
| 390 |
} |
| 391 |
|
| 392 |
ScmObj Scm__MethodDispatcherInfo(const ScmMethodDispatcher *dis) |
| 393 |
{ |
| 394 |
ScmObj h = SCM_NIL, t = SCM_NIL; |
| 395 |
/* Need to strip 'const', because of C11 error |
| 396 |
http://www.open-std.org/jtc1/sc22/wg14/www/docs/summary.htm#dr_459 */ |
| 397 |
ScmAtomicVar *loc = (ScmAtomicVar*)&dis->methodHash; |
| 398 |
const mhash *mh = (const mhash*)Scm_AtomicLoad(loc); |
| 399 |
SCM_APPEND1(h, t, SCM_MAKE_KEYWORD("axis")); |
| 400 |
SCM_APPEND1(h, t, SCM_MAKE_INT(dis->axis)); |
| 401 |
SCM_APPEND1(h, t, SCM_MAKE_KEYWORD("num-entries")); |
| 402 |
SCM_APPEND1(h, t, SCM_MAKE_INT(mh->num_entries)); |
| 403 |
return h; |
| 404 |
} |
| 405 |
|
| 406 |
void Scm__MethodDispatcherDump(ScmMethodDispatcher *dis, ScmPort *port) |
| 407 |
{ |
| 408 |
Scm_Printf(port, "MethodDispatcher axis=%d\n", dis->axis); |
| 409 |
mhash_print((mhash*)dis->methodHash, port); |
| 410 |
} |