1 | /* |
2 | * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. |
3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 | * |
5 | * This code is free software; you can redistribute it and/or modify it |
6 | * under the terms of the GNU General Public License version 2 only, as |
7 | * published by the Free Software Foundation. |
8 | * |
9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
12 | * version 2 for more details (a copy is included in the LICENSE file that |
13 | * accompanied this code). |
14 | * |
15 | * You should have received a copy of the GNU General Public License version |
16 | * 2 along with this work; if not, write to the Free Software Foundation, |
17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
18 | * |
19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
20 | * or visit www.oracle.com if you need additional information or have any |
21 | * questions. |
22 | * |
23 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "memory/allocation.inline.hpp" |
27 | #include "opto/addnode.hpp" |
28 | #include "opto/connode.hpp" |
29 | #include "opto/convertnode.hpp" |
30 | #include "opto/divnode.hpp" |
31 | #include "opto/machnode.hpp" |
32 | #include "opto/movenode.hpp" |
33 | #include "opto/matcher.hpp" |
34 | #include "opto/mulnode.hpp" |
35 | #include "opto/phaseX.hpp" |
36 | #include "opto/subnode.hpp" |
37 | |
38 | // Portions of code courtesy of Clifford Click |
39 | |
40 | // Optimization - Graph Style |
41 | |
42 | #include <math.h> |
43 | |
44 | //----------------------magic_int_divide_constants----------------------------- |
45 | // Compute magic multiplier and shift constant for converting a 32 bit divide |
46 | // by constant into a multiply/shift/add series. Return false if calculations |
47 | // fail. |
48 | // |
49 | // Borrowed almost verbatim from Hacker's Delight by Henry S. Warren, Jr. with |
50 | // minor type name and parameter changes. |
51 | static bool magic_int_divide_constants(jint d, jint &M, jint &s) { |
52 | int32_t p; |
53 | uint32_t ad, anc, delta, q1, r1, q2, r2, t; |
54 | const uint32_t two31 = 0x80000000L; // 2**31. |
55 | |
56 | ad = ABS(d); |
57 | if (d == 0 || d == 1) return false; |
58 | t = two31 + ((uint32_t)d >> 31); |
59 | anc = t - 1 - t%ad; // Absolute value of nc. |
60 | p = 31; // Init. p. |
61 | q1 = two31/anc; // Init. q1 = 2**p/|nc|. |
62 | r1 = two31 - q1*anc; // Init. r1 = rem(2**p, |nc|). |
63 | q2 = two31/ad; // Init. q2 = 2**p/|d|. |
64 | r2 = two31 - q2*ad; // Init. r2 = rem(2**p, |d|). |
65 | do { |
66 | p = p + 1; |
67 | q1 = 2*q1; // Update q1 = 2**p/|nc|. |
68 | r1 = 2*r1; // Update r1 = rem(2**p, |nc|). |
69 | if (r1 >= anc) { // (Must be an unsigned |
70 | q1 = q1 + 1; // comparison here). |
71 | r1 = r1 - anc; |
72 | } |
73 | q2 = 2*q2; // Update q2 = 2**p/|d|. |
74 | r2 = 2*r2; // Update r2 = rem(2**p, |d|). |
75 | if (r2 >= ad) { // (Must be an unsigned |
76 | q2 = q2 + 1; // comparison here). |
77 | r2 = r2 - ad; |
78 | } |
79 | delta = ad - r2; |
80 | } while (q1 < delta || (q1 == delta && r1 == 0)); |
81 | |
82 | M = q2 + 1; |
83 | if (d < 0) M = -M; // Magic number and |
84 | s = p - 32; // shift amount to return. |
85 | |
86 | return true; |
87 | } |
88 | |
89 | //--------------------------transform_int_divide------------------------------- |
90 | // Convert a division by constant divisor into an alternate Ideal graph. |
91 | // Return NULL if no transformation occurs. |
92 | static Node *transform_int_divide( PhaseGVN *phase, Node *dividend, jint divisor ) { |
93 | |
94 | // Check for invalid divisors |
95 | assert( divisor != 0 && divisor != min_jint, |
96 | "bad divisor for transforming to long multiply" ); |
97 | |
98 | bool d_pos = divisor >= 0; |
99 | jint d = d_pos ? divisor : -divisor; |
100 | const int N = 32; |
101 | |
102 | // Result |
103 | Node *q = NULL; |
104 | |
105 | if (d == 1) { |
106 | // division by +/- 1 |
107 | if (!d_pos) { |
108 | // Just negate the value |
109 | q = new SubINode(phase->intcon(0), dividend); |
110 | } |
111 | } else if ( is_power_of_2(d) ) { |
112 | // division by +/- a power of 2 |
113 | |
114 | // See if we can simply do a shift without rounding |
115 | bool needs_rounding = true; |
116 | const Type *dt = phase->type(dividend); |
117 | const TypeInt *dti = dt->isa_int(); |
118 | if (dti && dti->_lo >= 0) { |
119 | // we don't need to round a positive dividend |
120 | needs_rounding = false; |
121 | } else if( dividend->Opcode() == Op_AndI ) { |
122 | // An AND mask of sufficient size clears the low bits and |
123 | // I can avoid rounding. |
124 | const TypeInt *andconi_t = phase->type( dividend->in(2) )->isa_int(); |
125 | if( andconi_t && andconi_t->is_con() ) { |
126 | jint andconi = andconi_t->get_con(); |
127 | if( andconi < 0 && is_power_of_2(-andconi) && (-andconi) >= d ) { |
128 | if( (-andconi) == d ) // Remove AND if it clears bits which will be shifted |
129 | dividend = dividend->in(1); |
130 | needs_rounding = false; |
131 | } |
132 | } |
133 | } |
134 | |
135 | // Add rounding to the shift to handle the sign bit |
136 | int l = log2_jint(d-1)+1; |
137 | if (needs_rounding) { |
138 | // Divide-by-power-of-2 can be made into a shift, but you have to do |
139 | // more math for the rounding. You need to add 0 for positive |
140 | // numbers, and "i-1" for negative numbers. Example: i=4, so the |
141 | // shift is by 2. You need to add 3 to negative dividends and 0 to |
142 | // positive ones. So (-7+3)>>2 becomes -1, (-4+3)>>2 becomes -1, |
143 | // (-2+3)>>2 becomes 0, etc. |
144 | |
145 | // Compute 0 or -1, based on sign bit |
146 | Node *sign = phase->transform(new RShiftINode(dividend, phase->intcon(N - 1))); |
147 | // Mask sign bit to the low sign bits |
148 | Node *round = phase->transform(new URShiftINode(sign, phase->intcon(N - l))); |
149 | // Round up before shifting |
150 | dividend = phase->transform(new AddINode(dividend, round)); |
151 | } |
152 | |
153 | // Shift for division |
154 | q = new RShiftINode(dividend, phase->intcon(l)); |
155 | |
156 | if (!d_pos) { |
157 | q = new SubINode(phase->intcon(0), phase->transform(q)); |
158 | } |
159 | } else { |
160 | // Attempt the jint constant divide -> multiply transform found in |
161 | // "Division by Invariant Integers using Multiplication" |
162 | // by Granlund and Montgomery |
163 | // See also "Hacker's Delight", chapter 10 by Warren. |
164 | |
165 | jint magic_const; |
166 | jint shift_const; |
167 | if (magic_int_divide_constants(d, magic_const, shift_const)) { |
168 | Node *magic = phase->longcon(magic_const); |
169 | Node *dividend_long = phase->transform(new ConvI2LNode(dividend)); |
170 | |
171 | // Compute the high half of the dividend x magic multiplication |
172 | Node *mul_hi = phase->transform(new MulLNode(dividend_long, magic)); |
173 | |
174 | if (magic_const < 0) { |
175 | mul_hi = phase->transform(new RShiftLNode(mul_hi, phase->intcon(N))); |
176 | mul_hi = phase->transform(new ConvL2INode(mul_hi)); |
177 | |
178 | // The magic multiplier is too large for a 32 bit constant. We've adjusted |
179 | // it down by 2^32, but have to add 1 dividend back in after the multiplication. |
180 | // This handles the "overflow" case described by Granlund and Montgomery. |
181 | mul_hi = phase->transform(new AddINode(dividend, mul_hi)); |
182 | |
183 | // Shift over the (adjusted) mulhi |
184 | if (shift_const != 0) { |
185 | mul_hi = phase->transform(new RShiftINode(mul_hi, phase->intcon(shift_const))); |
186 | } |
187 | } else { |
188 | // No add is required, we can merge the shifts together. |
189 | mul_hi = phase->transform(new RShiftLNode(mul_hi, phase->intcon(N + shift_const))); |
190 | mul_hi = phase->transform(new ConvL2INode(mul_hi)); |
191 | } |
192 | |
193 | // Get a 0 or -1 from the sign of the dividend. |
194 | Node *addend0 = mul_hi; |
195 | Node *addend1 = phase->transform(new RShiftINode(dividend, phase->intcon(N-1))); |
196 | |
197 | // If the divisor is negative, swap the order of the input addends; |
198 | // this has the effect of negating the quotient. |
199 | if (!d_pos) { |
200 | Node *temp = addend0; addend0 = addend1; addend1 = temp; |
201 | } |
202 | |
203 | // Adjust the final quotient by subtracting -1 (adding 1) |
204 | // from the mul_hi. |
205 | q = new SubINode(addend0, addend1); |
206 | } |
207 | } |
208 | |
209 | return q; |
210 | } |
211 | |
212 | //---------------------magic_long_divide_constants----------------------------- |
213 | // Compute magic multiplier and shift constant for converting a 64 bit divide |
214 | // by constant into a multiply/shift/add series. Return false if calculations |
215 | // fail. |
216 | // |
217 | // Borrowed almost verbatim from Hacker's Delight by Henry S. Warren, Jr. with |
218 | // minor type name and parameter changes. Adjusted to 64 bit word width. |
219 | static bool magic_long_divide_constants(jlong d, jlong &M, jint &s) { |
220 | int64_t p; |
221 | uint64_t ad, anc, delta, q1, r1, q2, r2, t; |
222 | const uint64_t two63 = UCONST64(0x8000000000000000); // 2**63. |
223 | |
224 | ad = ABS(d); |
225 | if (d == 0 || d == 1) return false; |
226 | t = two63 + ((uint64_t)d >> 63); |
227 | anc = t - 1 - t%ad; // Absolute value of nc. |
228 | p = 63; // Init. p. |
229 | q1 = two63/anc; // Init. q1 = 2**p/|nc|. |
230 | r1 = two63 - q1*anc; // Init. r1 = rem(2**p, |nc|). |
231 | q2 = two63/ad; // Init. q2 = 2**p/|d|. |
232 | r2 = two63 - q2*ad; // Init. r2 = rem(2**p, |d|). |
233 | do { |
234 | p = p + 1; |
235 | q1 = 2*q1; // Update q1 = 2**p/|nc|. |
236 | r1 = 2*r1; // Update r1 = rem(2**p, |nc|). |
237 | if (r1 >= anc) { // (Must be an unsigned |
238 | q1 = q1 + 1; // comparison here). |
239 | r1 = r1 - anc; |
240 | } |
241 | q2 = 2*q2; // Update q2 = 2**p/|d|. |
242 | r2 = 2*r2; // Update r2 = rem(2**p, |d|). |
243 | if (r2 >= ad) { // (Must be an unsigned |
244 | q2 = q2 + 1; // comparison here). |
245 | r2 = r2 - ad; |
246 | } |
247 | delta = ad - r2; |
248 | } while (q1 < delta || (q1 == delta && r1 == 0)); |
249 | |
250 | M = q2 + 1; |
251 | if (d < 0) M = -M; // Magic number and |
252 | s = p - 64; // shift amount to return. |
253 | |
254 | return true; |
255 | } |
256 | |
257 | //---------------------long_by_long_mulhi-------------------------------------- |
258 | // Generate ideal node graph for upper half of a 64 bit x 64 bit multiplication |
259 | static Node* long_by_long_mulhi(PhaseGVN* phase, Node* dividend, jlong magic_const) { |
260 | // If the architecture supports a 64x64 mulhi, there is |
261 | // no need to synthesize it in ideal nodes. |
262 | if (Matcher::has_match_rule(Op_MulHiL)) { |
263 | Node* v = phase->longcon(magic_const); |
264 | return new MulHiLNode(dividend, v); |
265 | } |
266 | |
267 | // Taken from Hacker's Delight, Fig. 8-2. Multiply high signed. |
268 | // (http://www.hackersdelight.org/HDcode/mulhs.c) |
269 | // |
270 | // int mulhs(int u, int v) { |
271 | // unsigned u0, v0, w0; |
272 | // int u1, v1, w1, w2, t; |
273 | // |
274 | // u0 = u & 0xFFFF; u1 = u >> 16; |
275 | // v0 = v & 0xFFFF; v1 = v >> 16; |
276 | // w0 = u0*v0; |
277 | // t = u1*v0 + (w0 >> 16); |
278 | // w1 = t & 0xFFFF; |
279 | // w2 = t >> 16; |
280 | // w1 = u0*v1 + w1; |
281 | // return u1*v1 + w2 + (w1 >> 16); |
282 | // } |
283 | // |
284 | // Note: The version above is for 32x32 multiplications, while the |
285 | // following inline comments are adapted to 64x64. |
286 | |
287 | const int N = 64; |
288 | |
289 | // Dummy node to keep intermediate nodes alive during construction |
290 | Node* hook = new Node(4); |
291 | |
292 | // u0 = u & 0xFFFFFFFF; u1 = u >> 32; |
293 | Node* u0 = phase->transform(new AndLNode(dividend, phase->longcon(0xFFFFFFFF))); |
294 | Node* u1 = phase->transform(new RShiftLNode(dividend, phase->intcon(N / 2))); |
295 | hook->init_req(0, u0); |
296 | hook->init_req(1, u1); |
297 | |
298 | // v0 = v & 0xFFFFFFFF; v1 = v >> 32; |
299 | Node* v0 = phase->longcon(magic_const & 0xFFFFFFFF); |
300 | Node* v1 = phase->longcon(magic_const >> (N / 2)); |
301 | |
302 | // w0 = u0*v0; |
303 | Node* w0 = phase->transform(new MulLNode(u0, v0)); |
304 | |
305 | // t = u1*v0 + (w0 >> 32); |
306 | Node* u1v0 = phase->transform(new MulLNode(u1, v0)); |
307 | Node* temp = phase->transform(new URShiftLNode(w0, phase->intcon(N / 2))); |
308 | Node* t = phase->transform(new AddLNode(u1v0, temp)); |
309 | hook->init_req(2, t); |
310 | |
311 | // w1 = t & 0xFFFFFFFF; |
312 | Node* w1 = phase->transform(new AndLNode(t, phase->longcon(0xFFFFFFFF))); |
313 | hook->init_req(3, w1); |
314 | |
315 | // w2 = t >> 32; |
316 | Node* w2 = phase->transform(new RShiftLNode(t, phase->intcon(N / 2))); |
317 | |
318 | // w1 = u0*v1 + w1; |
319 | Node* u0v1 = phase->transform(new MulLNode(u0, v1)); |
320 | w1 = phase->transform(new AddLNode(u0v1, w1)); |
321 | |
322 | // return u1*v1 + w2 + (w1 >> 32); |
323 | Node* u1v1 = phase->transform(new MulLNode(u1, v1)); |
324 | Node* temp1 = phase->transform(new AddLNode(u1v1, w2)); |
325 | Node* temp2 = phase->transform(new RShiftLNode(w1, phase->intcon(N / 2))); |
326 | |
327 | // Remove the bogus extra edges used to keep things alive |
328 | PhaseIterGVN* igvn = phase->is_IterGVN(); |
329 | if (igvn != NULL) { |
330 | igvn->remove_dead_node(hook); |
331 | } else { |
332 | for (int i = 0; i < 4; i++) { |
333 | hook->set_req(i, NULL); |
334 | } |
335 | } |
336 | |
337 | return new AddLNode(temp1, temp2); |
338 | } |
339 | |
340 | |
341 | //--------------------------transform_long_divide------------------------------ |
342 | // Convert a division by constant divisor into an alternate Ideal graph. |
343 | // Return NULL if no transformation occurs. |
344 | static Node *transform_long_divide( PhaseGVN *phase, Node *dividend, jlong divisor ) { |
345 | // Check for invalid divisors |
346 | assert( divisor != 0L && divisor != min_jlong, |
347 | "bad divisor for transforming to long multiply" ); |
348 | |
349 | bool d_pos = divisor >= 0; |
350 | jlong d = d_pos ? divisor : -divisor; |
351 | const int N = 64; |
352 | |
353 | // Result |
354 | Node *q = NULL; |
355 | |
356 | if (d == 1) { |
357 | // division by +/- 1 |
358 | if (!d_pos) { |
359 | // Just negate the value |
360 | q = new SubLNode(phase->longcon(0), dividend); |
361 | } |
362 | } else if ( is_power_of_2_long(d) ) { |
363 | |
364 | // division by +/- a power of 2 |
365 | |
366 | // See if we can simply do a shift without rounding |
367 | bool needs_rounding = true; |
368 | const Type *dt = phase->type(dividend); |
369 | const TypeLong *dtl = dt->isa_long(); |
370 | |
371 | if (dtl && dtl->_lo > 0) { |
372 | // we don't need to round a positive dividend |
373 | needs_rounding = false; |
374 | } else if( dividend->Opcode() == Op_AndL ) { |
375 | // An AND mask of sufficient size clears the low bits and |
376 | // I can avoid rounding. |
377 | const TypeLong *andconl_t = phase->type( dividend->in(2) )->isa_long(); |
378 | if( andconl_t && andconl_t->is_con() ) { |
379 | jlong andconl = andconl_t->get_con(); |
380 | if( andconl < 0 && is_power_of_2_long(-andconl) && (-andconl) >= d ) { |
381 | if( (-andconl) == d ) // Remove AND if it clears bits which will be shifted |
382 | dividend = dividend->in(1); |
383 | needs_rounding = false; |
384 | } |
385 | } |
386 | } |
387 | |
388 | // Add rounding to the shift to handle the sign bit |
389 | int l = log2_long(d-1)+1; |
390 | if (needs_rounding) { |
391 | // Divide-by-power-of-2 can be made into a shift, but you have to do |
392 | // more math for the rounding. You need to add 0 for positive |
393 | // numbers, and "i-1" for negative numbers. Example: i=4, so the |
394 | // shift is by 2. You need to add 3 to negative dividends and 0 to |
395 | // positive ones. So (-7+3)>>2 becomes -1, (-4+3)>>2 becomes -1, |
396 | // (-2+3)>>2 becomes 0, etc. |
397 | |
398 | // Compute 0 or -1, based on sign bit |
399 | Node *sign = phase->transform(new RShiftLNode(dividend, phase->intcon(N - 1))); |
400 | // Mask sign bit to the low sign bits |
401 | Node *round = phase->transform(new URShiftLNode(sign, phase->intcon(N - l))); |
402 | // Round up before shifting |
403 | dividend = phase->transform(new AddLNode(dividend, round)); |
404 | } |
405 | |
406 | // Shift for division |
407 | q = new RShiftLNode(dividend, phase->intcon(l)); |
408 | |
409 | if (!d_pos) { |
410 | q = new SubLNode(phase->longcon(0), phase->transform(q)); |
411 | } |
412 | } else if ( !Matcher::use_asm_for_ldiv_by_con(d) ) { // Use hardware DIV instruction when |
413 | // it is faster than code generated below. |
414 | // Attempt the jlong constant divide -> multiply transform found in |
415 | // "Division by Invariant Integers using Multiplication" |
416 | // by Granlund and Montgomery |
417 | // See also "Hacker's Delight", chapter 10 by Warren. |
418 | |
419 | jlong magic_const; |
420 | jint shift_const; |
421 | if (magic_long_divide_constants(d, magic_const, shift_const)) { |
422 | // Compute the high half of the dividend x magic multiplication |
423 | Node *mul_hi = phase->transform(long_by_long_mulhi(phase, dividend, magic_const)); |
424 | |
425 | // The high half of the 128-bit multiply is computed. |
426 | if (magic_const < 0) { |
427 | // The magic multiplier is too large for a 64 bit constant. We've adjusted |
428 | // it down by 2^64, but have to add 1 dividend back in after the multiplication. |
429 | // This handles the "overflow" case described by Granlund and Montgomery. |
430 | mul_hi = phase->transform(new AddLNode(dividend, mul_hi)); |
431 | } |
432 | |
433 | // Shift over the (adjusted) mulhi |
434 | if (shift_const != 0) { |
435 | mul_hi = phase->transform(new RShiftLNode(mul_hi, phase->intcon(shift_const))); |
436 | } |
437 | |
438 | // Get a 0 or -1 from the sign of the dividend. |
439 | Node *addend0 = mul_hi; |
440 | Node *addend1 = phase->transform(new RShiftLNode(dividend, phase->intcon(N-1))); |
441 | |
442 | // If the divisor is negative, swap the order of the input addends; |
443 | // this has the effect of negating the quotient. |
444 | if (!d_pos) { |
445 | Node *temp = addend0; addend0 = addend1; addend1 = temp; |
446 | } |
447 | |
448 | // Adjust the final quotient by subtracting -1 (adding 1) |
449 | // from the mul_hi. |
450 | q = new SubLNode(addend0, addend1); |
451 | } |
452 | } |
453 | |
454 | return q; |
455 | } |
456 | |
457 | //============================================================================= |
458 | //------------------------------Identity--------------------------------------- |
459 | // If the divisor is 1, we are an identity on the dividend. |
460 | Node* DivINode::Identity(PhaseGVN* phase) { |
461 | return (phase->type( in(2) )->higher_equal(TypeInt::ONE)) ? in(1) : this; |
462 | } |
463 | |
464 | //------------------------------Idealize--------------------------------------- |
465 | // Divides can be changed to multiplies and/or shifts |
466 | Node *DivINode::Ideal(PhaseGVN *phase, bool can_reshape) { |
467 | if (in(0) && remove_dead_region(phase, can_reshape)) return this; |
468 | // Don't bother trying to transform a dead node |
469 | if( in(0) && in(0)->is_top() ) return NULL; |
470 | |
471 | const Type *t = phase->type( in(2) ); |
472 | if( t == TypeInt::ONE ) // Identity? |
473 | return NULL; // Skip it |
474 | |
475 | const TypeInt *ti = t->isa_int(); |
476 | if( !ti ) return NULL; |
477 | |
478 | // Check for useless control input |
479 | // Check for excluding div-zero case |
480 | if (in(0) && (ti->_hi < 0 || ti->_lo > 0)) { |
481 | set_req(0, NULL); // Yank control input |
482 | return this; |
483 | } |
484 | |
485 | if( !ti->is_con() ) return NULL; |
486 | jint i = ti->get_con(); // Get divisor |
487 | |
488 | if (i == 0) return NULL; // Dividing by zero constant does not idealize |
489 | |
490 | // Dividing by MININT does not optimize as a power-of-2 shift. |
491 | if( i == min_jint ) return NULL; |
492 | |
493 | return transform_int_divide( phase, in(1), i ); |
494 | } |
495 | |
496 | //------------------------------Value------------------------------------------ |
497 | // A DivINode divides its inputs. The third input is a Control input, used to |
498 | // prevent hoisting the divide above an unsafe test. |
499 | const Type* DivINode::Value(PhaseGVN* phase) const { |
500 | // Either input is TOP ==> the result is TOP |
501 | const Type *t1 = phase->type( in(1) ); |
502 | const Type *t2 = phase->type( in(2) ); |
503 | if( t1 == Type::TOP ) return Type::TOP; |
504 | if( t2 == Type::TOP ) return Type::TOP; |
505 | |
506 | // x/x == 1 since we always generate the dynamic divisor check for 0. |
507 | if( phase->eqv( in(1), in(2) ) ) |
508 | return TypeInt::ONE; |
509 | |
510 | // Either input is BOTTOM ==> the result is the local BOTTOM |
511 | const Type *bot = bottom_type(); |
512 | if( (t1 == bot) || (t2 == bot) || |
513 | (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) |
514 | return bot; |
515 | |
516 | // Divide the two numbers. We approximate. |
517 | // If divisor is a constant and not zero |
518 | const TypeInt *i1 = t1->is_int(); |
519 | const TypeInt *i2 = t2->is_int(); |
520 | int widen = MAX2(i1->_widen, i2->_widen); |
521 | |
522 | if( i2->is_con() && i2->get_con() != 0 ) { |
523 | int32_t d = i2->get_con(); // Divisor |
524 | jint lo, hi; |
525 | if( d >= 0 ) { |
526 | lo = i1->_lo/d; |
527 | hi = i1->_hi/d; |
528 | } else { |
529 | if( d == -1 && i1->_lo == min_jint ) { |
530 | // 'min_jint/-1' throws arithmetic exception during compilation |
531 | lo = min_jint; |
532 | // do not support holes, 'hi' must go to either min_jint or max_jint: |
533 | // [min_jint, -10]/[-1,-1] ==> [min_jint] UNION [10,max_jint] |
534 | hi = i1->_hi == min_jint ? min_jint : max_jint; |
535 | } else { |
536 | lo = i1->_hi/d; |
537 | hi = i1->_lo/d; |
538 | } |
539 | } |
540 | return TypeInt::make(lo, hi, widen); |
541 | } |
542 | |
543 | // If the dividend is a constant |
544 | if( i1->is_con() ) { |
545 | int32_t d = i1->get_con(); |
546 | if( d < 0 ) { |
547 | if( d == min_jint ) { |
548 | // (-min_jint) == min_jint == (min_jint / -1) |
549 | return TypeInt::make(min_jint, max_jint/2 + 1, widen); |
550 | } else { |
551 | return TypeInt::make(d, -d, widen); |
552 | } |
553 | } |
554 | return TypeInt::make(-d, d, widen); |
555 | } |
556 | |
557 | // Otherwise we give up all hope |
558 | return TypeInt::INT; |
559 | } |
560 | |
561 | |
562 | //============================================================================= |
563 | //------------------------------Identity--------------------------------------- |
564 | // If the divisor is 1, we are an identity on the dividend. |
565 | Node* DivLNode::Identity(PhaseGVN* phase) { |
566 | return (phase->type( in(2) )->higher_equal(TypeLong::ONE)) ? in(1) : this; |
567 | } |
568 | |
569 | //------------------------------Idealize--------------------------------------- |
570 | // Dividing by a power of 2 is a shift. |
571 | Node *DivLNode::Ideal( PhaseGVN *phase, bool can_reshape) { |
572 | if (in(0) && remove_dead_region(phase, can_reshape)) return this; |
573 | // Don't bother trying to transform a dead node |
574 | if( in(0) && in(0)->is_top() ) return NULL; |
575 | |
576 | const Type *t = phase->type( in(2) ); |
577 | if( t == TypeLong::ONE ) // Identity? |
578 | return NULL; // Skip it |
579 | |
580 | const TypeLong *tl = t->isa_long(); |
581 | if( !tl ) return NULL; |
582 | |
583 | // Check for useless control input |
584 | // Check for excluding div-zero case |
585 | if (in(0) && (tl->_hi < 0 || tl->_lo > 0)) { |
586 | set_req(0, NULL); // Yank control input |
587 | return this; |
588 | } |
589 | |
590 | if( !tl->is_con() ) return NULL; |
591 | jlong l = tl->get_con(); // Get divisor |
592 | |
593 | if (l == 0) return NULL; // Dividing by zero constant does not idealize |
594 | |
595 | // Dividing by MINLONG does not optimize as a power-of-2 shift. |
596 | if( l == min_jlong ) return NULL; |
597 | |
598 | return transform_long_divide( phase, in(1), l ); |
599 | } |
600 | |
601 | //------------------------------Value------------------------------------------ |
602 | // A DivLNode divides its inputs. The third input is a Control input, used to |
603 | // prevent hoisting the divide above an unsafe test. |
604 | const Type* DivLNode::Value(PhaseGVN* phase) const { |
605 | // Either input is TOP ==> the result is TOP |
606 | const Type *t1 = phase->type( in(1) ); |
607 | const Type *t2 = phase->type( in(2) ); |
608 | if( t1 == Type::TOP ) return Type::TOP; |
609 | if( t2 == Type::TOP ) return Type::TOP; |
610 | |
611 | // x/x == 1 since we always generate the dynamic divisor check for 0. |
612 | if( phase->eqv( in(1), in(2) ) ) |
613 | return TypeLong::ONE; |
614 | |
615 | // Either input is BOTTOM ==> the result is the local BOTTOM |
616 | const Type *bot = bottom_type(); |
617 | if( (t1 == bot) || (t2 == bot) || |
618 | (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) |
619 | return bot; |
620 | |
621 | // Divide the two numbers. We approximate. |
622 | // If divisor is a constant and not zero |
623 | const TypeLong *i1 = t1->is_long(); |
624 | const TypeLong *i2 = t2->is_long(); |
625 | int widen = MAX2(i1->_widen, i2->_widen); |
626 | |
627 | if( i2->is_con() && i2->get_con() != 0 ) { |
628 | jlong d = i2->get_con(); // Divisor |
629 | jlong lo, hi; |
630 | if( d >= 0 ) { |
631 | lo = i1->_lo/d; |
632 | hi = i1->_hi/d; |
633 | } else { |
634 | if( d == CONST64(-1) && i1->_lo == min_jlong ) { |
635 | // 'min_jlong/-1' throws arithmetic exception during compilation |
636 | lo = min_jlong; |
637 | // do not support holes, 'hi' must go to either min_jlong or max_jlong: |
638 | // [min_jlong, -10]/[-1,-1] ==> [min_jlong] UNION [10,max_jlong] |
639 | hi = i1->_hi == min_jlong ? min_jlong : max_jlong; |
640 | } else { |
641 | lo = i1->_hi/d; |
642 | hi = i1->_lo/d; |
643 | } |
644 | } |
645 | return TypeLong::make(lo, hi, widen); |
646 | } |
647 | |
648 | // If the dividend is a constant |
649 | if( i1->is_con() ) { |
650 | jlong d = i1->get_con(); |
651 | if( d < 0 ) { |
652 | if( d == min_jlong ) { |
653 | // (-min_jlong) == min_jlong == (min_jlong / -1) |
654 | return TypeLong::make(min_jlong, max_jlong/2 + 1, widen); |
655 | } else { |
656 | return TypeLong::make(d, -d, widen); |
657 | } |
658 | } |
659 | return TypeLong::make(-d, d, widen); |
660 | } |
661 | |
662 | // Otherwise we give up all hope |
663 | return TypeLong::LONG; |
664 | } |
665 | |
666 | |
667 | //============================================================================= |
668 | //------------------------------Value------------------------------------------ |
669 | // An DivFNode divides its inputs. The third input is a Control input, used to |
670 | // prevent hoisting the divide above an unsafe test. |
671 | const Type* DivFNode::Value(PhaseGVN* phase) const { |
672 | // Either input is TOP ==> the result is TOP |
673 | const Type *t1 = phase->type( in(1) ); |
674 | const Type *t2 = phase->type( in(2) ); |
675 | if( t1 == Type::TOP ) return Type::TOP; |
676 | if( t2 == Type::TOP ) return Type::TOP; |
677 | |
678 | // Either input is BOTTOM ==> the result is the local BOTTOM |
679 | const Type *bot = bottom_type(); |
680 | if( (t1 == bot) || (t2 == bot) || |
681 | (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) |
682 | return bot; |
683 | |
684 | // x/x == 1, we ignore 0/0. |
685 | // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) |
686 | // Does not work for variables because of NaN's |
687 | if( phase->eqv( in(1), in(2) ) && t1->base() == Type::FloatCon) |
688 | if (!g_isnan(t1->getf()) && g_isfinite(t1->getf()) && t1->getf() != 0.0) // could be negative ZERO or NaN |
689 | return TypeF::ONE; |
690 | |
691 | if( t2 == TypeF::ONE ) |
692 | return t1; |
693 | |
694 | // If divisor is a constant and not zero, divide them numbers |
695 | if( t1->base() == Type::FloatCon && |
696 | t2->base() == Type::FloatCon && |
697 | t2->getf() != 0.0 ) // could be negative zero |
698 | return TypeF::make( t1->getf()/t2->getf() ); |
699 | |
700 | // If the dividend is a constant zero |
701 | // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) |
702 | // Test TypeF::ZERO is not sufficient as it could be negative zero |
703 | |
704 | if( t1 == TypeF::ZERO && !g_isnan(t2->getf()) && t2->getf() != 0.0 ) |
705 | return TypeF::ZERO; |
706 | |
707 | // Otherwise we give up all hope |
708 | return Type::FLOAT; |
709 | } |
710 | |
711 | //------------------------------isA_Copy--------------------------------------- |
712 | // Dividing by self is 1. |
713 | // If the divisor is 1, we are an identity on the dividend. |
714 | Node* DivFNode::Identity(PhaseGVN* phase) { |
715 | return (phase->type( in(2) ) == TypeF::ONE) ? in(1) : this; |
716 | } |
717 | |
718 | |
719 | //------------------------------Idealize--------------------------------------- |
720 | Node *DivFNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
721 | if (in(0) && remove_dead_region(phase, can_reshape)) return this; |
722 | // Don't bother trying to transform a dead node |
723 | if( in(0) && in(0)->is_top() ) return NULL; |
724 | |
725 | const Type *t2 = phase->type( in(2) ); |
726 | if( t2 == TypeF::ONE ) // Identity? |
727 | return NULL; // Skip it |
728 | |
729 | const TypeF *tf = t2->isa_float_constant(); |
730 | if( !tf ) return NULL; |
731 | if( tf->base() != Type::FloatCon ) return NULL; |
732 | |
733 | // Check for out of range values |
734 | if( tf->is_nan() || !tf->is_finite() ) return NULL; |
735 | |
736 | // Get the value |
737 | float f = tf->getf(); |
738 | int exp; |
739 | |
740 | // Only for special case of dividing by a power of 2 |
741 | if( frexp((double)f, &exp) != 0.5 ) return NULL; |
742 | |
743 | // Limit the range of acceptable exponents |
744 | if( exp < -126 || exp > 126 ) return NULL; |
745 | |
746 | // Compute the reciprocal |
747 | float reciprocal = ((float)1.0) / f; |
748 | |
749 | assert( frexp((double)reciprocal, &exp) == 0.5, "reciprocal should be power of 2" ); |
750 | |
751 | // return multiplication by the reciprocal |
752 | return (new MulFNode(in(1), phase->makecon(TypeF::make(reciprocal)))); |
753 | } |
754 | |
755 | //============================================================================= |
756 | //------------------------------Value------------------------------------------ |
757 | // An DivDNode divides its inputs. The third input is a Control input, used to |
758 | // prevent hoisting the divide above an unsafe test. |
759 | const Type* DivDNode::Value(PhaseGVN* phase) const { |
760 | // Either input is TOP ==> the result is TOP |
761 | const Type *t1 = phase->type( in(1) ); |
762 | const Type *t2 = phase->type( in(2) ); |
763 | if( t1 == Type::TOP ) return Type::TOP; |
764 | if( t2 == Type::TOP ) return Type::TOP; |
765 | |
766 | // Either input is BOTTOM ==> the result is the local BOTTOM |
767 | const Type *bot = bottom_type(); |
768 | if( (t1 == bot) || (t2 == bot) || |
769 | (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) |
770 | return bot; |
771 | |
772 | // x/x == 1, we ignore 0/0. |
773 | // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) |
774 | // Does not work for variables because of NaN's |
775 | if( phase->eqv( in(1), in(2) ) && t1->base() == Type::DoubleCon) |
776 | if (!g_isnan(t1->getd()) && g_isfinite(t1->getd()) && t1->getd() != 0.0) // could be negative ZERO or NaN |
777 | return TypeD::ONE; |
778 | |
779 | if( t2 == TypeD::ONE ) |
780 | return t1; |
781 | |
782 | #if defined(IA32) |
783 | if (!phase->C->method()->is_strict()) |
784 | // Can't trust native compilers to properly fold strict double |
785 | // division with round-to-zero on this platform. |
786 | #endif |
787 | { |
788 | // If divisor is a constant and not zero, divide them numbers |
789 | if( t1->base() == Type::DoubleCon && |
790 | t2->base() == Type::DoubleCon && |
791 | t2->getd() != 0.0 ) // could be negative zero |
792 | return TypeD::make( t1->getd()/t2->getd() ); |
793 | } |
794 | |
795 | // If the dividend is a constant zero |
796 | // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) |
797 | // Test TypeF::ZERO is not sufficient as it could be negative zero |
798 | if( t1 == TypeD::ZERO && !g_isnan(t2->getd()) && t2->getd() != 0.0 ) |
799 | return TypeD::ZERO; |
800 | |
801 | // Otherwise we give up all hope |
802 | return Type::DOUBLE; |
803 | } |
804 | |
805 | |
806 | //------------------------------isA_Copy--------------------------------------- |
807 | // Dividing by self is 1. |
808 | // If the divisor is 1, we are an identity on the dividend. |
809 | Node* DivDNode::Identity(PhaseGVN* phase) { |
810 | return (phase->type( in(2) ) == TypeD::ONE) ? in(1) : this; |
811 | } |
812 | |
813 | //------------------------------Idealize--------------------------------------- |
814 | Node *DivDNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
815 | if (in(0) && remove_dead_region(phase, can_reshape)) return this; |
816 | // Don't bother trying to transform a dead node |
817 | if( in(0) && in(0)->is_top() ) return NULL; |
818 | |
819 | const Type *t2 = phase->type( in(2) ); |
820 | if( t2 == TypeD::ONE ) // Identity? |
821 | return NULL; // Skip it |
822 | |
823 | const TypeD *td = t2->isa_double_constant(); |
824 | if( !td ) return NULL; |
825 | if( td->base() != Type::DoubleCon ) return NULL; |
826 | |
827 | // Check for out of range values |
828 | if( td->is_nan() || !td->is_finite() ) return NULL; |
829 | |
830 | // Get the value |
831 | double d = td->getd(); |
832 | int exp; |
833 | |
834 | // Only for special case of dividing by a power of 2 |
835 | if( frexp(d, &exp) != 0.5 ) return NULL; |
836 | |
837 | // Limit the range of acceptable exponents |
838 | if( exp < -1021 || exp > 1022 ) return NULL; |
839 | |
840 | // Compute the reciprocal |
841 | double reciprocal = 1.0 / d; |
842 | |
843 | assert( frexp(reciprocal, &exp) == 0.5, "reciprocal should be power of 2" ); |
844 | |
845 | // return multiplication by the reciprocal |
846 | return (new MulDNode(in(1), phase->makecon(TypeD::make(reciprocal)))); |
847 | } |
848 | |
849 | //============================================================================= |
850 | //------------------------------Idealize--------------------------------------- |
851 | Node *ModINode::Ideal(PhaseGVN *phase, bool can_reshape) { |
852 | // Check for dead control input |
853 | if( in(0) && remove_dead_region(phase, can_reshape) ) return this; |
854 | // Don't bother trying to transform a dead node |
855 | if( in(0) && in(0)->is_top() ) return NULL; |
856 | |
857 | // Get the modulus |
858 | const Type *t = phase->type( in(2) ); |
859 | if( t == Type::TOP ) return NULL; |
860 | const TypeInt *ti = t->is_int(); |
861 | |
862 | // Check for useless control input |
863 | // Check for excluding mod-zero case |
864 | if (in(0) && (ti->_hi < 0 || ti->_lo > 0)) { |
865 | set_req(0, NULL); // Yank control input |
866 | return this; |
867 | } |
868 | |
869 | // See if we are MOD'ing by 2^k or 2^k-1. |
870 | if( !ti->is_con() ) return NULL; |
871 | jint con = ti->get_con(); |
872 | |
873 | Node *hook = new Node(1); |
874 | |
875 | // First, special check for modulo 2^k-1 |
876 | if( con >= 0 && con < max_jint && is_power_of_2(con+1) ) { |
877 | uint k = exact_log2(con+1); // Extract k |
878 | |
879 | // Basic algorithm by David Detlefs. See fastmod_int.java for gory details. |
880 | static int unroll_factor[] = { 999, 999, 29, 14, 9, 7, 5, 4, 4, 3, 3, 2, 2, 2, 2, 2, 1 /*past here we assume 1 forever*/}; |
881 | int trip_count = 1; |
882 | if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k]; |
883 | |
884 | // If the unroll factor is not too large, and if conditional moves are |
885 | // ok, then use this case |
886 | if( trip_count <= 5 && ConditionalMoveLimit != 0 ) { |
887 | Node *x = in(1); // Value being mod'd |
888 | Node *divisor = in(2); // Also is mask |
889 | |
890 | hook->init_req(0, x); // Add a use to x to prevent him from dying |
891 | // Generate code to reduce X rapidly to nearly 2^k-1. |
892 | for( int i = 0; i < trip_count; i++ ) { |
893 | Node *xl = phase->transform( new AndINode(x,divisor) ); |
894 | Node *xh = phase->transform( new RShiftINode(x,phase->intcon(k)) ); // Must be signed |
895 | x = phase->transform( new AddINode(xh,xl) ); |
896 | hook->set_req(0, x); |
897 | } |
898 | |
899 | // Generate sign-fixup code. Was original value positive? |
900 | // int hack_res = (i >= 0) ? divisor : 1; |
901 | Node *cmp1 = phase->transform( new CmpINode( in(1), phase->intcon(0) ) ); |
902 | Node *bol1 = phase->transform( new BoolNode( cmp1, BoolTest::ge ) ); |
903 | Node *cmov1= phase->transform( new CMoveINode(bol1, phase->intcon(1), divisor, TypeInt::POS) ); |
904 | // if( x >= hack_res ) x -= divisor; |
905 | Node *sub = phase->transform( new SubINode( x, divisor ) ); |
906 | Node *cmp2 = phase->transform( new CmpINode( x, cmov1 ) ); |
907 | Node *bol2 = phase->transform( new BoolNode( cmp2, BoolTest::ge ) ); |
908 | // Convention is to not transform the return value of an Ideal |
909 | // since Ideal is expected to return a modified 'this' or a new node. |
910 | Node *cmov2= new CMoveINode(bol2, x, sub, TypeInt::INT); |
911 | // cmov2 is now the mod |
912 | |
913 | // Now remove the bogus extra edges used to keep things alive |
914 | if (can_reshape) { |
915 | phase->is_IterGVN()->remove_dead_node(hook); |
916 | } else { |
917 | hook->set_req(0, NULL); // Just yank bogus edge during Parse phase |
918 | } |
919 | return cmov2; |
920 | } |
921 | } |
922 | |
923 | // Fell thru, the unroll case is not appropriate. Transform the modulo |
924 | // into a long multiply/int multiply/subtract case |
925 | |
926 | // Cannot handle mod 0, and min_jint isn't handled by the transform |
927 | if( con == 0 || con == min_jint ) return NULL; |
928 | |
929 | // Get the absolute value of the constant; at this point, we can use this |
930 | jint pos_con = (con >= 0) ? con : -con; |
931 | |
932 | // integer Mod 1 is always 0 |
933 | if( pos_con == 1 ) return new ConINode(TypeInt::ZERO); |
934 | |
935 | int log2_con = -1; |
936 | |
937 | // If this is a power of two, they maybe we can mask it |
938 | if( is_power_of_2(pos_con) ) { |
939 | log2_con = log2_intptr((intptr_t)pos_con); |
940 | |
941 | const Type *dt = phase->type(in(1)); |
942 | const TypeInt *dti = dt->isa_int(); |
943 | |
944 | // See if this can be masked, if the dividend is non-negative |
945 | if( dti && dti->_lo >= 0 ) |
946 | return ( new AndINode( in(1), phase->intcon( pos_con-1 ) ) ); |
947 | } |
948 | |
949 | // Save in(1) so that it cannot be changed or deleted |
950 | hook->init_req(0, in(1)); |
951 | |
952 | // Divide using the transform from DivI to MulL |
953 | Node *result = transform_int_divide( phase, in(1), pos_con ); |
954 | if (result != NULL) { |
955 | Node *divide = phase->transform(result); |
956 | |
957 | // Re-multiply, using a shift if this is a power of two |
958 | Node *mult = NULL; |
959 | |
960 | if( log2_con >= 0 ) |
961 | mult = phase->transform( new LShiftINode( divide, phase->intcon( log2_con ) ) ); |
962 | else |
963 | mult = phase->transform( new MulINode( divide, phase->intcon( pos_con ) ) ); |
964 | |
965 | // Finally, subtract the multiplied divided value from the original |
966 | result = new SubINode( in(1), mult ); |
967 | } |
968 | |
969 | // Now remove the bogus extra edges used to keep things alive |
970 | if (can_reshape) { |
971 | phase->is_IterGVN()->remove_dead_node(hook); |
972 | } else { |
973 | hook->set_req(0, NULL); // Just yank bogus edge during Parse phase |
974 | } |
975 | |
976 | // return the value |
977 | return result; |
978 | } |
979 | |
980 | //------------------------------Value------------------------------------------ |
981 | const Type* ModINode::Value(PhaseGVN* phase) const { |
982 | // Either input is TOP ==> the result is TOP |
983 | const Type *t1 = phase->type( in(1) ); |
984 | const Type *t2 = phase->type( in(2) ); |
985 | if( t1 == Type::TOP ) return Type::TOP; |
986 | if( t2 == Type::TOP ) return Type::TOP; |
987 | |
988 | // We always generate the dynamic check for 0. |
989 | // 0 MOD X is 0 |
990 | if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; |
991 | // X MOD X is 0 |
992 | if( phase->eqv( in(1), in(2) ) ) return TypeInt::ZERO; |
993 | |
994 | // Either input is BOTTOM ==> the result is the local BOTTOM |
995 | const Type *bot = bottom_type(); |
996 | if( (t1 == bot) || (t2 == bot) || |
997 | (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) |
998 | return bot; |
999 | |
1000 | const TypeInt *i1 = t1->is_int(); |
1001 | const TypeInt *i2 = t2->is_int(); |
1002 | if( !i1->is_con() || !i2->is_con() ) { |
1003 | if( i1->_lo >= 0 && i2->_lo >= 0 ) |
1004 | return TypeInt::POS; |
1005 | // If both numbers are not constants, we know little. |
1006 | return TypeInt::INT; |
1007 | } |
1008 | // Mod by zero? Throw exception at runtime! |
1009 | if( !i2->get_con() ) return TypeInt::POS; |
1010 | |
1011 | // We must be modulo'ing 2 float constants. |
1012 | // Check for min_jint % '-1', result is defined to be '0'. |
1013 | if( i1->get_con() == min_jint && i2->get_con() == -1 ) |
1014 | return TypeInt::ZERO; |
1015 | |
1016 | return TypeInt::make( i1->get_con() % i2->get_con() ); |
1017 | } |
1018 | |
1019 | |
1020 | //============================================================================= |
1021 | //------------------------------Idealize--------------------------------------- |
1022 | Node *ModLNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
1023 | // Check for dead control input |
1024 | if( in(0) && remove_dead_region(phase, can_reshape) ) return this; |
1025 | // Don't bother trying to transform a dead node |
1026 | if( in(0) && in(0)->is_top() ) return NULL; |
1027 | |
1028 | // Get the modulus |
1029 | const Type *t = phase->type( in(2) ); |
1030 | if( t == Type::TOP ) return NULL; |
1031 | const TypeLong *tl = t->is_long(); |
1032 | |
1033 | // Check for useless control input |
1034 | // Check for excluding mod-zero case |
1035 | if (in(0) && (tl->_hi < 0 || tl->_lo > 0)) { |
1036 | set_req(0, NULL); // Yank control input |
1037 | return this; |
1038 | } |
1039 | |
1040 | // See if we are MOD'ing by 2^k or 2^k-1. |
1041 | if( !tl->is_con() ) return NULL; |
1042 | jlong con = tl->get_con(); |
1043 | |
1044 | Node *hook = new Node(1); |
1045 | |
1046 | // Expand mod |
1047 | if( con >= 0 && con < max_jlong && is_power_of_2_long(con+1) ) { |
1048 | uint k = exact_log2_long(con+1); // Extract k |
1049 | |
1050 | // Basic algorithm by David Detlefs. See fastmod_long.java for gory details. |
1051 | // Used to help a popular random number generator which does a long-mod |
1052 | // of 2^31-1 and shows up in SpecJBB and SciMark. |
1053 | static int unroll_factor[] = { 999, 999, 61, 30, 20, 15, 12, 10, 8, 7, 6, 6, 5, 5, 4, 4, 4, 3, 3, 3, 3, 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1 /*past here we assume 1 forever*/}; |
1054 | int trip_count = 1; |
1055 | if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k]; |
1056 | |
1057 | // If the unroll factor is not too large, and if conditional moves are |
1058 | // ok, then use this case |
1059 | if( trip_count <= 5 && ConditionalMoveLimit != 0 ) { |
1060 | Node *x = in(1); // Value being mod'd |
1061 | Node *divisor = in(2); // Also is mask |
1062 | |
1063 | hook->init_req(0, x); // Add a use to x to prevent him from dying |
1064 | // Generate code to reduce X rapidly to nearly 2^k-1. |
1065 | for( int i = 0; i < trip_count; i++ ) { |
1066 | Node *xl = phase->transform( new AndLNode(x,divisor) ); |
1067 | Node *xh = phase->transform( new RShiftLNode(x,phase->intcon(k)) ); // Must be signed |
1068 | x = phase->transform( new AddLNode(xh,xl) ); |
1069 | hook->set_req(0, x); // Add a use to x to prevent him from dying |
1070 | } |
1071 | |
1072 | // Generate sign-fixup code. Was original value positive? |
1073 | // long hack_res = (i >= 0) ? divisor : CONST64(1); |
1074 | Node *cmp1 = phase->transform( new CmpLNode( in(1), phase->longcon(0) ) ); |
1075 | Node *bol1 = phase->transform( new BoolNode( cmp1, BoolTest::ge ) ); |
1076 | Node *cmov1= phase->transform( new CMoveLNode(bol1, phase->longcon(1), divisor, TypeLong::LONG) ); |
1077 | // if( x >= hack_res ) x -= divisor; |
1078 | Node *sub = phase->transform( new SubLNode( x, divisor ) ); |
1079 | Node *cmp2 = phase->transform( new CmpLNode( x, cmov1 ) ); |
1080 | Node *bol2 = phase->transform( new BoolNode( cmp2, BoolTest::ge ) ); |
1081 | // Convention is to not transform the return value of an Ideal |
1082 | // since Ideal is expected to return a modified 'this' or a new node. |
1083 | Node *cmov2= new CMoveLNode(bol2, x, sub, TypeLong::LONG); |
1084 | // cmov2 is now the mod |
1085 | |
1086 | // Now remove the bogus extra edges used to keep things alive |
1087 | if (can_reshape) { |
1088 | phase->is_IterGVN()->remove_dead_node(hook); |
1089 | } else { |
1090 | hook->set_req(0, NULL); // Just yank bogus edge during Parse phase |
1091 | } |
1092 | return cmov2; |
1093 | } |
1094 | } |
1095 | |
1096 | // Fell thru, the unroll case is not appropriate. Transform the modulo |
1097 | // into a long multiply/int multiply/subtract case |
1098 | |
1099 | // Cannot handle mod 0, and min_jlong isn't handled by the transform |
1100 | if( con == 0 || con == min_jlong ) return NULL; |
1101 | |
1102 | // Get the absolute value of the constant; at this point, we can use this |
1103 | jlong pos_con = (con >= 0) ? con : -con; |
1104 | |
1105 | // integer Mod 1 is always 0 |
1106 | if( pos_con == 1 ) return new ConLNode(TypeLong::ZERO); |
1107 | |
1108 | int log2_con = -1; |
1109 | |
1110 | // If this is a power of two, then maybe we can mask it |
1111 | if( is_power_of_2_long(pos_con) ) { |
1112 | log2_con = exact_log2_long(pos_con); |
1113 | |
1114 | const Type *dt = phase->type(in(1)); |
1115 | const TypeLong *dtl = dt->isa_long(); |
1116 | |
1117 | // See if this can be masked, if the dividend is non-negative |
1118 | if( dtl && dtl->_lo >= 0 ) |
1119 | return ( new AndLNode( in(1), phase->longcon( pos_con-1 ) ) ); |
1120 | } |
1121 | |
1122 | // Save in(1) so that it cannot be changed or deleted |
1123 | hook->init_req(0, in(1)); |
1124 | |
1125 | // Divide using the transform from DivL to MulL |
1126 | Node *result = transform_long_divide( phase, in(1), pos_con ); |
1127 | if (result != NULL) { |
1128 | Node *divide = phase->transform(result); |
1129 | |
1130 | // Re-multiply, using a shift if this is a power of two |
1131 | Node *mult = NULL; |
1132 | |
1133 | if( log2_con >= 0 ) |
1134 | mult = phase->transform( new LShiftLNode( divide, phase->intcon( log2_con ) ) ); |
1135 | else |
1136 | mult = phase->transform( new MulLNode( divide, phase->longcon( pos_con ) ) ); |
1137 | |
1138 | // Finally, subtract the multiplied divided value from the original |
1139 | result = new SubLNode( in(1), mult ); |
1140 | } |
1141 | |
1142 | // Now remove the bogus extra edges used to keep things alive |
1143 | if (can_reshape) { |
1144 | phase->is_IterGVN()->remove_dead_node(hook); |
1145 | } else { |
1146 | hook->set_req(0, NULL); // Just yank bogus edge during Parse phase |
1147 | } |
1148 | |
1149 | // return the value |
1150 | return result; |
1151 | } |
1152 | |
1153 | //------------------------------Value------------------------------------------ |
1154 | const Type* ModLNode::Value(PhaseGVN* phase) const { |
1155 | // Either input is TOP ==> the result is TOP |
1156 | const Type *t1 = phase->type( in(1) ); |
1157 | const Type *t2 = phase->type( in(2) ); |
1158 | if( t1 == Type::TOP ) return Type::TOP; |
1159 | if( t2 == Type::TOP ) return Type::TOP; |
1160 | |
1161 | // We always generate the dynamic check for 0. |
1162 | // 0 MOD X is 0 |
1163 | if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; |
1164 | // X MOD X is 0 |
1165 | if( phase->eqv( in(1), in(2) ) ) return TypeLong::ZERO; |
1166 | |
1167 | // Either input is BOTTOM ==> the result is the local BOTTOM |
1168 | const Type *bot = bottom_type(); |
1169 | if( (t1 == bot) || (t2 == bot) || |
1170 | (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) |
1171 | return bot; |
1172 | |
1173 | const TypeLong *i1 = t1->is_long(); |
1174 | const TypeLong *i2 = t2->is_long(); |
1175 | if( !i1->is_con() || !i2->is_con() ) { |
1176 | if( i1->_lo >= CONST64(0) && i2->_lo >= CONST64(0) ) |
1177 | return TypeLong::POS; |
1178 | // If both numbers are not constants, we know little. |
1179 | return TypeLong::LONG; |
1180 | } |
1181 | // Mod by zero? Throw exception at runtime! |
1182 | if( !i2->get_con() ) return TypeLong::POS; |
1183 | |
1184 | // We must be modulo'ing 2 float constants. |
1185 | // Check for min_jint % '-1', result is defined to be '0'. |
1186 | if( i1->get_con() == min_jlong && i2->get_con() == -1 ) |
1187 | return TypeLong::ZERO; |
1188 | |
1189 | return TypeLong::make( i1->get_con() % i2->get_con() ); |
1190 | } |
1191 | |
1192 | |
1193 | //============================================================================= |
1194 | //------------------------------Value------------------------------------------ |
1195 | const Type* ModFNode::Value(PhaseGVN* phase) const { |
1196 | // Either input is TOP ==> the result is TOP |
1197 | const Type *t1 = phase->type( in(1) ); |
1198 | const Type *t2 = phase->type( in(2) ); |
1199 | if( t1 == Type::TOP ) return Type::TOP; |
1200 | if( t2 == Type::TOP ) return Type::TOP; |
1201 | |
1202 | // Either input is BOTTOM ==> the result is the local BOTTOM |
1203 | const Type *bot = bottom_type(); |
1204 | if( (t1 == bot) || (t2 == bot) || |
1205 | (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) |
1206 | return bot; |
1207 | |
1208 | // If either number is not a constant, we know nothing. |
1209 | if ((t1->base() != Type::FloatCon) || (t2->base() != Type::FloatCon)) { |
1210 | return Type::FLOAT; // note: x%x can be either NaN or 0 |
1211 | } |
1212 | |
1213 | float f1 = t1->getf(); |
1214 | float f2 = t2->getf(); |
1215 | jint x1 = jint_cast(f1); // note: *(int*)&f1, not just (int)f1 |
1216 | jint x2 = jint_cast(f2); |
1217 | |
1218 | // If either is a NaN, return an input NaN |
1219 | if (g_isnan(f1)) return t1; |
1220 | if (g_isnan(f2)) return t2; |
1221 | |
1222 | // If an operand is infinity or the divisor is +/- zero, punt. |
1223 | if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jint) |
1224 | return Type::FLOAT; |
1225 | |
1226 | // We must be modulo'ing 2 float constants. |
1227 | // Make sure that the sign of the fmod is equal to the sign of the dividend |
1228 | jint xr = jint_cast(fmod(f1, f2)); |
1229 | if ((x1 ^ xr) < 0) { |
1230 | xr ^= min_jint; |
1231 | } |
1232 | |
1233 | return TypeF::make(jfloat_cast(xr)); |
1234 | } |
1235 | |
1236 | |
1237 | //============================================================================= |
1238 | //------------------------------Value------------------------------------------ |
1239 | const Type* ModDNode::Value(PhaseGVN* phase) const { |
1240 | // Either input is TOP ==> the result is TOP |
1241 | const Type *t1 = phase->type( in(1) ); |
1242 | const Type *t2 = phase->type( in(2) ); |
1243 | if( t1 == Type::TOP ) return Type::TOP; |
1244 | if( t2 == Type::TOP ) return Type::TOP; |
1245 | |
1246 | // Either input is BOTTOM ==> the result is the local BOTTOM |
1247 | const Type *bot = bottom_type(); |
1248 | if( (t1 == bot) || (t2 == bot) || |
1249 | (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) |
1250 | return bot; |
1251 | |
1252 | // If either number is not a constant, we know nothing. |
1253 | if ((t1->base() != Type::DoubleCon) || (t2->base() != Type::DoubleCon)) { |
1254 | return Type::DOUBLE; // note: x%x can be either NaN or 0 |
1255 | } |
1256 | |
1257 | double f1 = t1->getd(); |
1258 | double f2 = t2->getd(); |
1259 | jlong x1 = jlong_cast(f1); // note: *(long*)&f1, not just (long)f1 |
1260 | jlong x2 = jlong_cast(f2); |
1261 | |
1262 | // If either is a NaN, return an input NaN |
1263 | if (g_isnan(f1)) return t1; |
1264 | if (g_isnan(f2)) return t2; |
1265 | |
1266 | // If an operand is infinity or the divisor is +/- zero, punt. |
1267 | if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jlong) |
1268 | return Type::DOUBLE; |
1269 | |
1270 | // We must be modulo'ing 2 double constants. |
1271 | // Make sure that the sign of the fmod is equal to the sign of the dividend |
1272 | jlong xr = jlong_cast(fmod(f1, f2)); |
1273 | if ((x1 ^ xr) < 0) { |
1274 | xr ^= min_jlong; |
1275 | } |
1276 | |
1277 | return TypeD::make(jdouble_cast(xr)); |
1278 | } |
1279 | |
1280 | //============================================================================= |
1281 | |
1282 | DivModNode::DivModNode( Node *c, Node *dividend, Node *divisor ) : MultiNode(3) { |
1283 | init_req(0, c); |
1284 | init_req(1, dividend); |
1285 | init_req(2, divisor); |
1286 | } |
1287 | |
1288 | //------------------------------make------------------------------------------ |
1289 | DivModINode* DivModINode::make(Node* div_or_mod) { |
1290 | Node* n = div_or_mod; |
1291 | assert(n->Opcode() == Op_DivI || n->Opcode() == Op_ModI, |
1292 | "only div or mod input pattern accepted" ); |
1293 | |
1294 | DivModINode* divmod = new DivModINode(n->in(0), n->in(1), n->in(2)); |
1295 | Node* dproj = new ProjNode(divmod, DivModNode::div_proj_num); |
1296 | Node* mproj = new ProjNode(divmod, DivModNode::mod_proj_num); |
1297 | return divmod; |
1298 | } |
1299 | |
1300 | //------------------------------make------------------------------------------ |
1301 | DivModLNode* DivModLNode::make(Node* div_or_mod) { |
1302 | Node* n = div_or_mod; |
1303 | assert(n->Opcode() == Op_DivL || n->Opcode() == Op_ModL, |
1304 | "only div or mod input pattern accepted" ); |
1305 | |
1306 | DivModLNode* divmod = new DivModLNode(n->in(0), n->in(1), n->in(2)); |
1307 | Node* dproj = new ProjNode(divmod, DivModNode::div_proj_num); |
1308 | Node* mproj = new ProjNode(divmod, DivModNode::mod_proj_num); |
1309 | return divmod; |
1310 | } |
1311 | |
1312 | //------------------------------match------------------------------------------ |
1313 | // return result(s) along with their RegMask info |
1314 | Node *DivModINode::match( const ProjNode *proj, const Matcher *match ) { |
1315 | uint ideal_reg = proj->ideal_reg(); |
1316 | RegMask rm; |
1317 | if (proj->_con == div_proj_num) { |
1318 | rm = match->divI_proj_mask(); |
1319 | } else { |
1320 | assert(proj->_con == mod_proj_num, "must be div or mod projection" ); |
1321 | rm = match->modI_proj_mask(); |
1322 | } |
1323 | return new MachProjNode(this, proj->_con, rm, ideal_reg); |
1324 | } |
1325 | |
1326 | |
1327 | //------------------------------match------------------------------------------ |
1328 | // return result(s) along with their RegMask info |
1329 | Node *DivModLNode::match( const ProjNode *proj, const Matcher *match ) { |
1330 | uint ideal_reg = proj->ideal_reg(); |
1331 | RegMask rm; |
1332 | if (proj->_con == div_proj_num) { |
1333 | rm = match->divL_proj_mask(); |
1334 | } else { |
1335 | assert(proj->_con == mod_proj_num, "must be div or mod projection" ); |
1336 | rm = match->modL_proj_mask(); |
1337 | } |
1338 | return new MachProjNode(this, proj->_con, rm, ideal_reg); |
1339 | } |
1340 | |