Mercurial > hg > truffle
annotate src/share/vm/opto/divnode.cpp @ 145:f3de1255b035
6603011: RFE: Optimize long division
Summary: Transform long division by constant into multiply
Reviewed-by: never, kvn
author | rasbold |
---|---|
date | Wed, 07 May 2008 08:06:46 -0700 |
parents | 6e825ad773c6 |
children | d1605aabd0a1 |
rev | line source |
---|---|
0 | 1 /* |
2 * Copyright 1997-2006 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
21 * have any questions. | |
22 * | |
23 */ | |
24 | |
25 // Portions of code courtesy of Clifford Click | |
26 | |
27 // Optimization - Graph Style | |
28 | |
29 #include "incls/_precompiled.incl" | |
30 #include "incls/_divnode.cpp.incl" | |
31 #include <math.h> | |
32 | |
145 | 33 //----------------------magic_int_divide_constants----------------------------- |
34 // Compute magic multiplier and shift constant for converting a 32 bit divide | |
35 // by constant into a multiply/shift/add series. Return false if calculations | |
36 // fail. | |
37 // | |
38 // Borrowed almost verbatum from Hacker's Delight by Henry S. Warren, Jr. with | |
39 // minor type name and parameter changes. | |
40 static bool magic_int_divide_constants(jint d, jint &M, jint &s) { | |
41 int32_t p; | |
42 uint32_t ad, anc, delta, q1, r1, q2, r2, t; | |
43 const uint32_t two31 = 0x80000000L; // 2**31. | |
44 | |
45 ad = ABS(d); | |
46 if (d == 0 || d == 1) return false; | |
47 t = two31 + ((uint32_t)d >> 31); | |
48 anc = t - 1 - t%ad; // Absolute value of nc. | |
49 p = 31; // Init. p. | |
50 q1 = two31/anc; // Init. q1 = 2**p/|nc|. | |
51 r1 = two31 - q1*anc; // Init. r1 = rem(2**p, |nc|). | |
52 q2 = two31/ad; // Init. q2 = 2**p/|d|. | |
53 r2 = two31 - q2*ad; // Init. r2 = rem(2**p, |d|). | |
54 do { | |
55 p = p + 1; | |
56 q1 = 2*q1; // Update q1 = 2**p/|nc|. | |
57 r1 = 2*r1; // Update r1 = rem(2**p, |nc|). | |
58 if (r1 >= anc) { // (Must be an unsigned | |
59 q1 = q1 + 1; // comparison here). | |
60 r1 = r1 - anc; | |
61 } | |
62 q2 = 2*q2; // Update q2 = 2**p/|d|. | |
63 r2 = 2*r2; // Update r2 = rem(2**p, |d|). | |
64 if (r2 >= ad) { // (Must be an unsigned | |
65 q2 = q2 + 1; // comparison here). | |
66 r2 = r2 - ad; | |
67 } | |
68 delta = ad - r2; | |
69 } while (q1 < delta || (q1 == delta && r1 == 0)); | |
70 | |
71 M = q2 + 1; | |
72 if (d < 0) M = -M; // Magic number and | |
73 s = p - 32; // shift amount to return. | |
74 | |
75 return true; | |
76 } | |
77 | |
78 //--------------------------transform_int_divide------------------------------- | |
79 // Convert a division by constant divisor into an alternate Ideal graph. | |
80 // Return NULL if no transformation occurs. | |
81 static Node *transform_int_divide( PhaseGVN *phase, Node *dividend, jint divisor ) { | |
0 | 82 |
83 // Check for invalid divisors | |
145 | 84 assert( divisor != 0 && divisor != min_jint, |
85 "bad divisor for transforming to long multiply" ); | |
0 | 86 |
145 | 87 bool d_pos = divisor >= 0; |
88 jint d = d_pos ? divisor : -divisor; | |
89 const int N = 32; | |
0 | 90 |
91 // Result | |
145 | 92 Node *q = NULL; |
0 | 93 |
94 if (d == 1) { | |
145 | 95 // division by +/- 1 |
96 if (!d_pos) { | |
97 // Just negate the value | |
0 | 98 q = new (phase->C, 3) SubINode(phase->intcon(0), dividend); |
99 } | |
145 | 100 } else if ( is_power_of_2(d) ) { |
101 // division by +/- a power of 2 | |
0 | 102 |
103 // See if we can simply do a shift without rounding | |
104 bool needs_rounding = true; | |
105 const Type *dt = phase->type(dividend); | |
106 const TypeInt *dti = dt->isa_int(); | |
145 | 107 if (dti && dti->_lo >= 0) { |
108 // we don't need to round a positive dividend | |
0 | 109 needs_rounding = false; |
145 | 110 } else if( dividend->Opcode() == Op_AndI ) { |
111 // An AND mask of sufficient size clears the low bits and | |
112 // I can avoid rounding. | |
0 | 113 const TypeInt *andconi = phase->type( dividend->in(2) )->isa_int(); |
114 if( andconi && andconi->is_con(-d) ) { | |
115 dividend = dividend->in(1); | |
116 needs_rounding = false; | |
117 } | |
118 } | |
119 | |
120 // Add rounding to the shift to handle the sign bit | |
145 | 121 int l = log2_intptr(d-1)+1; |
122 if (needs_rounding) { | |
123 // Divide-by-power-of-2 can be made into a shift, but you have to do | |
124 // more math for the rounding. You need to add 0 for positive | |
125 // numbers, and "i-1" for negative numbers. Example: i=4, so the | |
126 // shift is by 2. You need to add 3 to negative dividends and 0 to | |
127 // positive ones. So (-7+3)>>2 becomes -1, (-4+3)>>2 becomes -1, | |
128 // (-2+3)>>2 becomes 0, etc. | |
129 | |
130 // Compute 0 or -1, based on sign bit | |
131 Node *sign = phase->transform(new (phase->C, 3) RShiftINode(dividend, phase->intcon(N - 1))); | |
132 // Mask sign bit to the low sign bits | |
133 Node *round = phase->transform(new (phase->C, 3) URShiftINode(sign, phase->intcon(N - l))); | |
134 // Round up before shifting | |
135 dividend = phase->transform(new (phase->C, 3) AddINode(dividend, round)); | |
0 | 136 } |
137 | |
145 | 138 // Shift for division |
0 | 139 q = new (phase->C, 3) RShiftINode(dividend, phase->intcon(l)); |
140 | |
145 | 141 if (!d_pos) { |
0 | 142 q = new (phase->C, 3) SubINode(phase->intcon(0), phase->transform(q)); |
145 | 143 } |
144 } else { | |
145 // Attempt the jint constant divide -> multiply transform found in | |
146 // "Division by Invariant Integers using Multiplication" | |
147 // by Granlund and Montgomery | |
148 // See also "Hacker's Delight", chapter 10 by Warren. | |
149 | |
150 jint magic_const; | |
151 jint shift_const; | |
152 if (magic_int_divide_constants(d, magic_const, shift_const)) { | |
153 Node *magic = phase->longcon(magic_const); | |
154 Node *dividend_long = phase->transform(new (phase->C, 2) ConvI2LNode(dividend)); | |
155 | |
156 // Compute the high half of the dividend x magic multiplication | |
157 Node *mul_hi = phase->transform(new (phase->C, 3) MulLNode(dividend_long, magic)); | |
158 | |
159 if (magic_const < 0) { | |
160 mul_hi = phase->transform(new (phase->C, 3) RShiftLNode(mul_hi, phase->intcon(N))); | |
161 mul_hi = phase->transform(new (phase->C, 2) ConvL2INode(mul_hi)); | |
162 | |
163 // The magic multiplier is too large for a 32 bit constant. We've adjusted | |
164 // it down by 2^32, but have to add 1 dividend back in after the multiplication. | |
165 // This handles the "overflow" case described by Granlund and Montgomery. | |
166 mul_hi = phase->transform(new (phase->C, 3) AddINode(dividend, mul_hi)); | |
167 | |
168 // Shift over the (adjusted) mulhi | |
169 if (shift_const != 0) { | |
170 mul_hi = phase->transform(new (phase->C, 3) RShiftINode(mul_hi, phase->intcon(shift_const))); | |
171 } | |
172 } else { | |
173 // No add is required, we can merge the shifts together. | |
174 mul_hi = phase->transform(new (phase->C, 3) RShiftLNode(mul_hi, phase->intcon(N + shift_const))); | |
175 mul_hi = phase->transform(new (phase->C, 2) ConvL2INode(mul_hi)); | |
176 } | |
177 | |
178 // Get a 0 or -1 from the sign of the dividend. | |
179 Node *addend0 = mul_hi; | |
180 Node *addend1 = phase->transform(new (phase->C, 3) RShiftINode(dividend, phase->intcon(N-1))); | |
181 | |
182 // If the divisor is negative, swap the order of the input addends; | |
183 // this has the effect of negating the quotient. | |
184 if (!d_pos) { | |
185 Node *temp = addend0; addend0 = addend1; addend1 = temp; | |
186 } | |
187 | |
188 // Adjust the final quotient by subtracting -1 (adding 1) | |
189 // from the mul_hi. | |
190 q = new (phase->C, 3) SubINode(addend0, addend1); | |
191 } | |
192 } | |
193 | |
194 return q; | |
195 } | |
196 | |
197 //---------------------magic_long_divide_constants----------------------------- | |
198 // Compute magic multiplier and shift constant for converting a 64 bit divide | |
199 // by constant into a multiply/shift/add series. Return false if calculations | |
200 // fail. | |
201 // | |
202 // Borrowed almost verbatum from Hacker's Delight by Henry S. Warren, Jr. with | |
203 // minor type name and parameter changes. Adjusted to 64 bit word width. | |
204 static bool magic_long_divide_constants(jlong d, jlong &M, jint &s) { | |
205 int64_t p; | |
206 uint64_t ad, anc, delta, q1, r1, q2, r2, t; | |
207 const uint64_t two63 = 0x8000000000000000LL; // 2**63. | |
208 | |
209 ad = ABS(d); | |
210 if (d == 0 || d == 1) return false; | |
211 t = two63 + ((uint64_t)d >> 63); | |
212 anc = t - 1 - t%ad; // Absolute value of nc. | |
213 p = 63; // Init. p. | |
214 q1 = two63/anc; // Init. q1 = 2**p/|nc|. | |
215 r1 = two63 - q1*anc; // Init. r1 = rem(2**p, |nc|). | |
216 q2 = two63/ad; // Init. q2 = 2**p/|d|. | |
217 r2 = two63 - q2*ad; // Init. r2 = rem(2**p, |d|). | |
218 do { | |
219 p = p + 1; | |
220 q1 = 2*q1; // Update q1 = 2**p/|nc|. | |
221 r1 = 2*r1; // Update r1 = rem(2**p, |nc|). | |
222 if (r1 >= anc) { // (Must be an unsigned | |
223 q1 = q1 + 1; // comparison here). | |
224 r1 = r1 - anc; | |
225 } | |
226 q2 = 2*q2; // Update q2 = 2**p/|d|. | |
227 r2 = 2*r2; // Update r2 = rem(2**p, |d|). | |
228 if (r2 >= ad) { // (Must be an unsigned | |
229 q2 = q2 + 1; // comparison here). | |
230 r2 = r2 - ad; | |
231 } | |
232 delta = ad - r2; | |
233 } while (q1 < delta || (q1 == delta && r1 == 0)); | |
234 | |
235 M = q2 + 1; | |
236 if (d < 0) M = -M; // Magic number and | |
237 s = p - 64; // shift amount to return. | |
238 | |
239 return true; | |
240 } | |
241 | |
242 //---------------------long_by_long_mulhi-------------------------------------- | |
243 // Generate ideal node graph for upper half of a 64 bit x 64 bit multiplication | |
244 static Node *long_by_long_mulhi( PhaseGVN *phase, Node *dividend, jlong magic_const) { | |
245 // If the architecture supports a 64x64 mulhi, there is | |
246 // no need to synthesize it in ideal nodes. | |
247 if (Matcher::has_match_rule(Op_MulHiL)) { | |
248 Node *v = phase->longcon(magic_const); | |
249 return new (phase->C, 3) MulHiLNode(dividend, v); | |
0 | 250 } |
251 | |
145 | 252 const int N = 64; |
253 | |
254 Node *u_hi = phase->transform(new (phase->C, 3) RShiftLNode(dividend, phase->intcon(N / 2))); | |
255 Node *u_lo = phase->transform(new (phase->C, 3) AndLNode(dividend, phase->longcon(0xFFFFFFFF))); | |
256 | |
257 Node *v_hi = phase->longcon(magic_const >> N/2); | |
258 Node *v_lo = phase->longcon(magic_const & 0XFFFFFFFF); | |
259 | |
260 Node *hihi_product = phase->transform(new (phase->C, 3) MulLNode(u_hi, v_hi)); | |
261 Node *hilo_product = phase->transform(new (phase->C, 3) MulLNode(u_hi, v_lo)); | |
262 Node *lohi_product = phase->transform(new (phase->C, 3) MulLNode(u_lo, v_hi)); | |
263 Node *lolo_product = phase->transform(new (phase->C, 3) MulLNode(u_lo, v_lo)); | |
264 | |
265 Node *t1 = phase->transform(new (phase->C, 3) URShiftLNode(lolo_product, phase->intcon(N / 2))); | |
266 Node *t2 = phase->transform(new (phase->C, 3) AddLNode(hilo_product, t1)); | |
267 Node *t3 = phase->transform(new (phase->C, 3) RShiftLNode(t2, phase->intcon(N / 2))); | |
268 Node *t4 = phase->transform(new (phase->C, 3) AndLNode(t2, phase->longcon(0xFFFFFFFF))); | |
269 Node *t5 = phase->transform(new (phase->C, 3) AddLNode(t4, lohi_product)); | |
270 Node *t6 = phase->transform(new (phase->C, 3) RShiftLNode(t5, phase->intcon(N / 2))); | |
271 Node *t7 = phase->transform(new (phase->C, 3) AddLNode(t3, hihi_product)); | |
272 | |
273 return new (phase->C, 3) AddLNode(t7, t6); | |
274 } | |
275 | |
276 | |
277 //--------------------------transform_long_divide------------------------------ | |
278 // Convert a division by constant divisor into an alternate Ideal graph. | |
279 // Return NULL if no transformation occurs. | |
280 static Node *transform_long_divide( PhaseGVN *phase, Node *dividend, jlong divisor ) { | |
281 // Check for invalid divisors | |
282 assert( divisor != 0L && divisor != min_jlong, | |
283 "bad divisor for transforming to long multiply" ); | |
284 | |
285 bool d_pos = divisor >= 0; | |
286 jlong d = d_pos ? divisor : -divisor; | |
287 const int N = 64; | |
288 | |
289 // Result | |
290 Node *q = NULL; | |
291 | |
292 if (d == 1) { | |
293 // division by +/- 1 | |
294 if (!d_pos) { | |
295 // Just negate the value | |
296 q = new (phase->C, 3) SubLNode(phase->longcon(0), dividend); | |
297 } | |
298 } else if ( is_power_of_2_long(d) ) { | |
299 | |
300 // division by +/- a power of 2 | |
301 | |
302 // See if we can simply do a shift without rounding | |
303 bool needs_rounding = true; | |
304 const Type *dt = phase->type(dividend); | |
305 const TypeLong *dtl = dt->isa_long(); | |
0 | 306 |
145 | 307 if (dtl && dtl->_lo > 0) { |
308 // we don't need to round a positive dividend | |
309 needs_rounding = false; | |
310 } else if( dividend->Opcode() == Op_AndL ) { | |
311 // An AND mask of sufficient size clears the low bits and | |
312 // I can avoid rounding. | |
313 const TypeLong *andconl = phase->type( dividend->in(2) )->isa_long(); | |
314 if( andconl && andconl->is_con(-d)) { | |
315 dividend = dividend->in(1); | |
316 needs_rounding = false; | |
317 } | |
318 } | |
319 | |
320 // Add rounding to the shift to handle the sign bit | |
321 int l = log2_long(d-1)+1; | |
322 if (needs_rounding) { | |
323 // Divide-by-power-of-2 can be made into a shift, but you have to do | |
324 // more math for the rounding. You need to add 0 for positive | |
325 // numbers, and "i-1" for negative numbers. Example: i=4, so the | |
326 // shift is by 2. You need to add 3 to negative dividends and 0 to | |
327 // positive ones. So (-7+3)>>2 becomes -1, (-4+3)>>2 becomes -1, | |
328 // (-2+3)>>2 becomes 0, etc. | |
329 | |
330 // Compute 0 or -1, based on sign bit | |
331 Node *sign = phase->transform(new (phase->C, 3) RShiftLNode(dividend, phase->intcon(N - 1))); | |
332 // Mask sign bit to the low sign bits | |
333 Node *round = phase->transform(new (phase->C, 3) URShiftLNode(sign, phase->intcon(N - l))); | |
334 // Round up before shifting | |
335 dividend = phase->transform(new (phase->C, 3) AddLNode(dividend, round)); | |
336 } | |
337 | |
338 // Shift for division | |
339 q = new (phase->C, 3) RShiftLNode(dividend, phase->intcon(l)); | |
340 | |
341 if (!d_pos) { | |
342 q = new (phase->C, 3) SubLNode(phase->longcon(0), phase->transform(q)); | |
343 } | |
344 } else { | |
345 // Attempt the jlong constant divide -> multiply transform found in | |
346 // "Division by Invariant Integers using Multiplication" | |
347 // by Granlund and Montgomery | |
348 // See also "Hacker's Delight", chapter 10 by Warren. | |
349 | |
350 jlong magic_const; | |
351 jint shift_const; | |
352 if (magic_long_divide_constants(d, magic_const, shift_const)) { | |
353 // Compute the high half of the dividend x magic multiplication | |
354 Node *mul_hi = phase->transform(long_by_long_mulhi(phase, dividend, magic_const)); | |
355 | |
356 // The high half of the 128-bit multiply is computed. | |
357 if (magic_const < 0) { | |
358 // The magic multiplier is too large for a 64 bit constant. We've adjusted | |
359 // it down by 2^64, but have to add 1 dividend back in after the multiplication. | |
360 // This handles the "overflow" case described by Granlund and Montgomery. | |
361 mul_hi = phase->transform(new (phase->C, 3) AddLNode(dividend, mul_hi)); | |
362 } | |
363 | |
364 // Shift over the (adjusted) mulhi | |
365 if (shift_const != 0) { | |
366 mul_hi = phase->transform(new (phase->C, 3) RShiftLNode(mul_hi, phase->intcon(shift_const))); | |
367 } | |
368 | |
369 // Get a 0 or -1 from the sign of the dividend. | |
370 Node *addend0 = mul_hi; | |
371 Node *addend1 = phase->transform(new (phase->C, 3) RShiftLNode(dividend, phase->intcon(N-1))); | |
372 | |
373 // If the divisor is negative, swap the order of the input addends; | |
374 // this has the effect of negating the quotient. | |
375 if (!d_pos) { | |
376 Node *temp = addend0; addend0 = addend1; addend1 = temp; | |
377 } | |
378 | |
379 // Adjust the final quotient by subtracting -1 (adding 1) | |
380 // from the mul_hi. | |
381 q = new (phase->C, 3) SubLNode(addend0, addend1); | |
382 } | |
0 | 383 } |
384 | |
145 | 385 return q; |
0 | 386 } |
387 | |
388 //============================================================================= | |
389 //------------------------------Identity--------------------------------------- | |
390 // If the divisor is 1, we are an identity on the dividend. | |
391 Node *DivINode::Identity( PhaseTransform *phase ) { | |
392 return (phase->type( in(2) )->higher_equal(TypeInt::ONE)) ? in(1) : this; | |
393 } | |
394 | |
395 //------------------------------Idealize--------------------------------------- | |
396 // Divides can be changed to multiplies and/or shifts | |
397 Node *DivINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
398 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
399 | |
400 const Type *t = phase->type( in(2) ); | |
401 if( t == TypeInt::ONE ) // Identity? | |
402 return NULL; // Skip it | |
403 | |
404 const TypeInt *ti = t->isa_int(); | |
405 if( !ti ) return NULL; | |
406 if( !ti->is_con() ) return NULL; | |
145 | 407 jint i = ti->get_con(); // Get divisor |
0 | 408 |
409 if (i == 0) return NULL; // Dividing by zero constant does not idealize | |
410 | |
411 set_req(0,NULL); // Dividing by a not-zero constant; no faulting | |
412 | |
413 // Dividing by MININT does not optimize as a power-of-2 shift. | |
414 if( i == min_jint ) return NULL; | |
415 | |
145 | 416 return transform_int_divide( phase, in(1), i ); |
0 | 417 } |
418 | |
419 //------------------------------Value------------------------------------------ | |
420 // A DivINode divides its inputs. The third input is a Control input, used to | |
421 // prevent hoisting the divide above an unsafe test. | |
422 const Type *DivINode::Value( PhaseTransform *phase ) const { | |
423 // Either input is TOP ==> the result is TOP | |
424 const Type *t1 = phase->type( in(1) ); | |
425 const Type *t2 = phase->type( in(2) ); | |
426 if( t1 == Type::TOP ) return Type::TOP; | |
427 if( t2 == Type::TOP ) return Type::TOP; | |
428 | |
429 // x/x == 1 since we always generate the dynamic divisor check for 0. | |
430 if( phase->eqv( in(1), in(2) ) ) | |
431 return TypeInt::ONE; | |
432 | |
433 // Either input is BOTTOM ==> the result is the local BOTTOM | |
434 const Type *bot = bottom_type(); | |
435 if( (t1 == bot) || (t2 == bot) || | |
436 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
437 return bot; | |
438 | |
439 // Divide the two numbers. We approximate. | |
440 // If divisor is a constant and not zero | |
441 const TypeInt *i1 = t1->is_int(); | |
442 const TypeInt *i2 = t2->is_int(); | |
443 int widen = MAX2(i1->_widen, i2->_widen); | |
444 | |
445 if( i2->is_con() && i2->get_con() != 0 ) { | |
446 int32 d = i2->get_con(); // Divisor | |
447 jint lo, hi; | |
448 if( d >= 0 ) { | |
449 lo = i1->_lo/d; | |
450 hi = i1->_hi/d; | |
451 } else { | |
452 if( d == -1 && i1->_lo == min_jint ) { | |
453 // 'min_jint/-1' throws arithmetic exception during compilation | |
454 lo = min_jint; | |
455 // do not support holes, 'hi' must go to either min_jint or max_jint: | |
456 // [min_jint, -10]/[-1,-1] ==> [min_jint] UNION [10,max_jint] | |
457 hi = i1->_hi == min_jint ? min_jint : max_jint; | |
458 } else { | |
459 lo = i1->_hi/d; | |
460 hi = i1->_lo/d; | |
461 } | |
462 } | |
463 return TypeInt::make(lo, hi, widen); | |
464 } | |
465 | |
466 // If the dividend is a constant | |
467 if( i1->is_con() ) { | |
468 int32 d = i1->get_con(); | |
469 if( d < 0 ) { | |
470 if( d == min_jint ) { | |
471 // (-min_jint) == min_jint == (min_jint / -1) | |
472 return TypeInt::make(min_jint, max_jint/2 + 1, widen); | |
473 } else { | |
474 return TypeInt::make(d, -d, widen); | |
475 } | |
476 } | |
477 return TypeInt::make(-d, d, widen); | |
478 } | |
479 | |
480 // Otherwise we give up all hope | |
481 return TypeInt::INT; | |
482 } | |
483 | |
484 | |
485 //============================================================================= | |
486 //------------------------------Identity--------------------------------------- | |
487 // If the divisor is 1, we are an identity on the dividend. | |
488 Node *DivLNode::Identity( PhaseTransform *phase ) { | |
489 return (phase->type( in(2) )->higher_equal(TypeLong::ONE)) ? in(1) : this; | |
490 } | |
491 | |
492 //------------------------------Idealize--------------------------------------- | |
493 // Dividing by a power of 2 is a shift. | |
494 Node *DivLNode::Ideal( PhaseGVN *phase, bool can_reshape) { | |
495 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
496 | |
497 const Type *t = phase->type( in(2) ); | |
145 | 498 if( t == TypeLong::ONE ) // Identity? |
0 | 499 return NULL; // Skip it |
500 | |
145 | 501 const TypeLong *tl = t->isa_long(); |
502 if( !tl ) return NULL; | |
503 if( !tl->is_con() ) return NULL; | |
504 jlong l = tl->get_con(); // Get divisor | |
505 | |
506 if (l == 0) return NULL; // Dividing by zero constant does not idealize | |
507 | |
508 set_req(0,NULL); // Dividing by a not-zero constant; no faulting | |
0 | 509 |
510 // Dividing by MININT does not optimize as a power-of-2 shift. | |
145 | 511 if( l == min_jlong ) return NULL; |
0 | 512 |
145 | 513 return transform_long_divide( phase, in(1), l ); |
0 | 514 } |
515 | |
516 //------------------------------Value------------------------------------------ | |
517 // A DivLNode divides its inputs. The third input is a Control input, used to | |
518 // prevent hoisting the divide above an unsafe test. | |
519 const Type *DivLNode::Value( PhaseTransform *phase ) const { | |
520 // Either input is TOP ==> the result is TOP | |
521 const Type *t1 = phase->type( in(1) ); | |
522 const Type *t2 = phase->type( in(2) ); | |
523 if( t1 == Type::TOP ) return Type::TOP; | |
524 if( t2 == Type::TOP ) return Type::TOP; | |
525 | |
526 // x/x == 1 since we always generate the dynamic divisor check for 0. | |
527 if( phase->eqv( in(1), in(2) ) ) | |
528 return TypeLong::ONE; | |
529 | |
530 // Either input is BOTTOM ==> the result is the local BOTTOM | |
531 const Type *bot = bottom_type(); | |
532 if( (t1 == bot) || (t2 == bot) || | |
533 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
534 return bot; | |
535 | |
536 // Divide the two numbers. We approximate. | |
537 // If divisor is a constant and not zero | |
538 const TypeLong *i1 = t1->is_long(); | |
539 const TypeLong *i2 = t2->is_long(); | |
540 int widen = MAX2(i1->_widen, i2->_widen); | |
541 | |
542 if( i2->is_con() && i2->get_con() != 0 ) { | |
543 jlong d = i2->get_con(); // Divisor | |
544 jlong lo, hi; | |
545 if( d >= 0 ) { | |
546 lo = i1->_lo/d; | |
547 hi = i1->_hi/d; | |
548 } else { | |
549 if( d == CONST64(-1) && i1->_lo == min_jlong ) { | |
550 // 'min_jlong/-1' throws arithmetic exception during compilation | |
551 lo = min_jlong; | |
552 // do not support holes, 'hi' must go to either min_jlong or max_jlong: | |
553 // [min_jlong, -10]/[-1,-1] ==> [min_jlong] UNION [10,max_jlong] | |
554 hi = i1->_hi == min_jlong ? min_jlong : max_jlong; | |
555 } else { | |
556 lo = i1->_hi/d; | |
557 hi = i1->_lo/d; | |
558 } | |
559 } | |
560 return TypeLong::make(lo, hi, widen); | |
561 } | |
562 | |
563 // If the dividend is a constant | |
564 if( i1->is_con() ) { | |
565 jlong d = i1->get_con(); | |
566 if( d < 0 ) { | |
567 if( d == min_jlong ) { | |
568 // (-min_jlong) == min_jlong == (min_jlong / -1) | |
569 return TypeLong::make(min_jlong, max_jlong/2 + 1, widen); | |
570 } else { | |
571 return TypeLong::make(d, -d, widen); | |
572 } | |
573 } | |
574 return TypeLong::make(-d, d, widen); | |
575 } | |
576 | |
577 // Otherwise we give up all hope | |
578 return TypeLong::LONG; | |
579 } | |
580 | |
581 | |
582 //============================================================================= | |
583 //------------------------------Value------------------------------------------ | |
584 // An DivFNode divides its inputs. The third input is a Control input, used to | |
585 // prevent hoisting the divide above an unsafe test. | |
586 const Type *DivFNode::Value( PhaseTransform *phase ) const { | |
587 // Either input is TOP ==> the result is TOP | |
588 const Type *t1 = phase->type( in(1) ); | |
589 const Type *t2 = phase->type( in(2) ); | |
590 if( t1 == Type::TOP ) return Type::TOP; | |
591 if( t2 == Type::TOP ) return Type::TOP; | |
592 | |
593 // Either input is BOTTOM ==> the result is the local BOTTOM | |
594 const Type *bot = bottom_type(); | |
595 if( (t1 == bot) || (t2 == bot) || | |
596 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
597 return bot; | |
598 | |
599 // x/x == 1, we ignore 0/0. | |
600 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
601 // Does not work for variables because of NaN's |
0 | 602 if( phase->eqv( in(1), in(2) ) && t1->base() == Type::FloatCon) |
603 if (!g_isnan(t1->getf()) && g_isfinite(t1->getf()) && t1->getf() != 0.0) // could be negative ZERO or NaN | |
604 return TypeF::ONE; | |
605 | |
606 if( t2 == TypeF::ONE ) | |
607 return t1; | |
608 | |
609 // If divisor is a constant and not zero, divide them numbers | |
610 if( t1->base() == Type::FloatCon && | |
611 t2->base() == Type::FloatCon && | |
612 t2->getf() != 0.0 ) // could be negative zero | |
613 return TypeF::make( t1->getf()/t2->getf() ); | |
614 | |
615 // If the dividend is a constant zero | |
616 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
617 // Test TypeF::ZERO is not sufficient as it could be negative zero | |
618 | |
619 if( t1 == TypeF::ZERO && !g_isnan(t2->getf()) && t2->getf() != 0.0 ) | |
620 return TypeF::ZERO; | |
621 | |
622 // Otherwise we give up all hope | |
623 return Type::FLOAT; | |
624 } | |
625 | |
626 //------------------------------isA_Copy--------------------------------------- | |
627 // Dividing by self is 1. | |
628 // If the divisor is 1, we are an identity on the dividend. | |
629 Node *DivFNode::Identity( PhaseTransform *phase ) { | |
630 return (phase->type( in(2) ) == TypeF::ONE) ? in(1) : this; | |
631 } | |
632 | |
633 | |
634 //------------------------------Idealize--------------------------------------- | |
635 Node *DivFNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
636 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
637 | |
638 const Type *t2 = phase->type( in(2) ); | |
639 if( t2 == TypeF::ONE ) // Identity? | |
640 return NULL; // Skip it | |
641 | |
642 const TypeF *tf = t2->isa_float_constant(); | |
643 if( !tf ) return NULL; | |
644 if( tf->base() != Type::FloatCon ) return NULL; | |
645 | |
646 // Check for out of range values | |
647 if( tf->is_nan() || !tf->is_finite() ) return NULL; | |
648 | |
649 // Get the value | |
650 float f = tf->getf(); | |
651 int exp; | |
652 | |
653 // Only for special case of dividing by a power of 2 | |
654 if( frexp((double)f, &exp) != 0.5 ) return NULL; | |
655 | |
656 // Limit the range of acceptable exponents | |
657 if( exp < -126 || exp > 126 ) return NULL; | |
658 | |
659 // Compute the reciprocal | |
660 float reciprocal = ((float)1.0) / f; | |
661 | |
662 assert( frexp((double)reciprocal, &exp) == 0.5, "reciprocal should be power of 2" ); | |
663 | |
664 // return multiplication by the reciprocal | |
665 return (new (phase->C, 3) MulFNode(in(1), phase->makecon(TypeF::make(reciprocal)))); | |
666 } | |
667 | |
668 //============================================================================= | |
669 //------------------------------Value------------------------------------------ | |
670 // An DivDNode divides its inputs. The third input is a Control input, used to | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
671 // prevent hoisting the divide above an unsafe test. |
0 | 672 const Type *DivDNode::Value( PhaseTransform *phase ) const { |
673 // Either input is TOP ==> the result is TOP | |
674 const Type *t1 = phase->type( in(1) ); | |
675 const Type *t2 = phase->type( in(2) ); | |
676 if( t1 == Type::TOP ) return Type::TOP; | |
677 if( t2 == Type::TOP ) return Type::TOP; | |
678 | |
679 // Either input is BOTTOM ==> the result is the local BOTTOM | |
680 const Type *bot = bottom_type(); | |
681 if( (t1 == bot) || (t2 == bot) || | |
682 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
683 return bot; | |
684 | |
685 // x/x == 1, we ignore 0/0. | |
686 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
687 // Does not work for variables because of NaN's | |
688 if( phase->eqv( in(1), in(2) ) && t1->base() == Type::DoubleCon) | |
689 if (!g_isnan(t1->getd()) && g_isfinite(t1->getd()) && t1->getd() != 0.0) // could be negative ZERO or NaN | |
690 return TypeD::ONE; | |
691 | |
692 if( t2 == TypeD::ONE ) | |
693 return t1; | |
694 | |
695 // If divisor is a constant and not zero, divide them numbers | |
696 if( t1->base() == Type::DoubleCon && | |
697 t2->base() == Type::DoubleCon && | |
698 t2->getd() != 0.0 ) // could be negative zero | |
699 return TypeD::make( t1->getd()/t2->getd() ); | |
700 | |
701 // If the dividend is a constant zero | |
702 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
703 // Test TypeF::ZERO is not sufficient as it could be negative zero | |
704 if( t1 == TypeD::ZERO && !g_isnan(t2->getd()) && t2->getd() != 0.0 ) | |
705 return TypeD::ZERO; | |
706 | |
707 // Otherwise we give up all hope | |
708 return Type::DOUBLE; | |
709 } | |
710 | |
711 | |
712 //------------------------------isA_Copy--------------------------------------- | |
713 // Dividing by self is 1. | |
714 // If the divisor is 1, we are an identity on the dividend. | |
715 Node *DivDNode::Identity( PhaseTransform *phase ) { | |
716 return (phase->type( in(2) ) == TypeD::ONE) ? in(1) : this; | |
717 } | |
718 | |
719 //------------------------------Idealize--------------------------------------- | |
720 Node *DivDNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
721 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
722 | |
723 const Type *t2 = phase->type( in(2) ); | |
724 if( t2 == TypeD::ONE ) // Identity? | |
725 return NULL; // Skip it | |
726 | |
727 const TypeD *td = t2->isa_double_constant(); | |
728 if( !td ) return NULL; | |
729 if( td->base() != Type::DoubleCon ) return NULL; | |
730 | |
731 // Check for out of range values | |
732 if( td->is_nan() || !td->is_finite() ) return NULL; | |
733 | |
734 // Get the value | |
735 double d = td->getd(); | |
736 int exp; | |
737 | |
738 // Only for special case of dividing by a power of 2 | |
739 if( frexp(d, &exp) != 0.5 ) return NULL; | |
740 | |
741 // Limit the range of acceptable exponents | |
742 if( exp < -1021 || exp > 1022 ) return NULL; | |
743 | |
744 // Compute the reciprocal | |
745 double reciprocal = 1.0 / d; | |
746 | |
747 assert( frexp(reciprocal, &exp) == 0.5, "reciprocal should be power of 2" ); | |
748 | |
749 // return multiplication by the reciprocal | |
750 return (new (phase->C, 3) MulDNode(in(1), phase->makecon(TypeD::make(reciprocal)))); | |
751 } | |
752 | |
753 //============================================================================= | |
754 //------------------------------Idealize--------------------------------------- | |
755 Node *ModINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
756 // Check for dead control input | |
757 if( remove_dead_region(phase, can_reshape) ) return this; | |
758 | |
759 // Get the modulus | |
760 const Type *t = phase->type( in(2) ); | |
761 if( t == Type::TOP ) return NULL; | |
762 const TypeInt *ti = t->is_int(); | |
763 | |
764 // Check for useless control input | |
765 // Check for excluding mod-zero case | |
766 if( in(0) && (ti->_hi < 0 || ti->_lo > 0) ) { | |
767 set_req(0, NULL); // Yank control input | |
768 return this; | |
769 } | |
770 | |
771 // See if we are MOD'ing by 2^k or 2^k-1. | |
772 if( !ti->is_con() ) return NULL; | |
773 jint con = ti->get_con(); | |
774 | |
775 Node *hook = new (phase->C, 1) Node(1); | |
776 | |
777 // First, special check for modulo 2^k-1 | |
778 if( con >= 0 && con < max_jint && is_power_of_2(con+1) ) { | |
779 uint k = exact_log2(con+1); // Extract k | |
780 | |
781 // Basic algorithm by David Detlefs. See fastmod_int.java for gory details. | |
782 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*/}; | |
783 int trip_count = 1; | |
784 if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k]; | |
785 | |
786 // If the unroll factor is not too large, and if conditional moves are | |
787 // ok, then use this case | |
788 if( trip_count <= 5 && ConditionalMoveLimit != 0 ) { | |
789 Node *x = in(1); // Value being mod'd | |
790 Node *divisor = in(2); // Also is mask | |
791 | |
792 hook->init_req(0, x); // Add a use to x to prevent him from dying | |
793 // Generate code to reduce X rapidly to nearly 2^k-1. | |
794 for( int i = 0; i < trip_count; i++ ) { | |
145 | 795 Node *xl = phase->transform( new (phase->C, 3) AndINode(x,divisor) ); |
796 Node *xh = phase->transform( new (phase->C, 3) RShiftINode(x,phase->intcon(k)) ); // Must be signed | |
797 x = phase->transform( new (phase->C, 3) AddINode(xh,xl) ); | |
798 hook->set_req(0, x); | |
0 | 799 } |
800 | |
801 // Generate sign-fixup code. Was original value positive? | |
802 // int hack_res = (i >= 0) ? divisor : 1; | |
803 Node *cmp1 = phase->transform( new (phase->C, 3) CmpINode( in(1), phase->intcon(0) ) ); | |
804 Node *bol1 = phase->transform( new (phase->C, 2) BoolNode( cmp1, BoolTest::ge ) ); | |
805 Node *cmov1= phase->transform( new (phase->C, 4) CMoveINode(bol1, phase->intcon(1), divisor, TypeInt::POS) ); | |
806 // if( x >= hack_res ) x -= divisor; | |
807 Node *sub = phase->transform( new (phase->C, 3) SubINode( x, divisor ) ); | |
808 Node *cmp2 = phase->transform( new (phase->C, 3) CmpINode( x, cmov1 ) ); | |
809 Node *bol2 = phase->transform( new (phase->C, 2) BoolNode( cmp2, BoolTest::ge ) ); | |
810 // Convention is to not transform the return value of an Ideal | |
811 // since Ideal is expected to return a modified 'this' or a new node. | |
812 Node *cmov2= new (phase->C, 4) CMoveINode(bol2, x, sub, TypeInt::INT); | |
813 // cmov2 is now the mod | |
814 | |
815 // Now remove the bogus extra edges used to keep things alive | |
816 if (can_reshape) { | |
817 phase->is_IterGVN()->remove_dead_node(hook); | |
818 } else { | |
819 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
820 } | |
821 return cmov2; | |
822 } | |
823 } | |
824 | |
825 // Fell thru, the unroll case is not appropriate. Transform the modulo | |
826 // into a long multiply/int multiply/subtract case | |
827 | |
828 // Cannot handle mod 0, and min_jint isn't handled by the transform | |
829 if( con == 0 || con == min_jint ) return NULL; | |
830 | |
831 // Get the absolute value of the constant; at this point, we can use this | |
832 jint pos_con = (con >= 0) ? con : -con; | |
833 | |
834 // integer Mod 1 is always 0 | |
835 if( pos_con == 1 ) return new (phase->C, 1) ConINode(TypeInt::ZERO); | |
836 | |
837 int log2_con = -1; | |
838 | |
839 // If this is a power of two, they maybe we can mask it | |
840 if( is_power_of_2(pos_con) ) { | |
841 log2_con = log2_intptr((intptr_t)pos_con); | |
842 | |
843 const Type *dt = phase->type(in(1)); | |
844 const TypeInt *dti = dt->isa_int(); | |
845 | |
846 // See if this can be masked, if the dividend is non-negative | |
847 if( dti && dti->_lo >= 0 ) | |
848 return ( new (phase->C, 3) AndINode( in(1), phase->intcon( pos_con-1 ) ) ); | |
849 } | |
850 | |
851 // Save in(1) so that it cannot be changed or deleted | |
852 hook->init_req(0, in(1)); | |
853 | |
854 // Divide using the transform from DivI to MulL | |
145 | 855 Node *result = transform_int_divide( phase, in(1), pos_con ); |
856 if (result != NULL) { | |
857 Node *divide = phase->transform(result); | |
0 | 858 |
145 | 859 // Re-multiply, using a shift if this is a power of two |
860 Node *mult = NULL; | |
0 | 861 |
145 | 862 if( log2_con >= 0 ) |
863 mult = phase->transform( new (phase->C, 3) LShiftINode( divide, phase->intcon( log2_con ) ) ); | |
864 else | |
865 mult = phase->transform( new (phase->C, 3) MulINode( divide, phase->intcon( pos_con ) ) ); | |
0 | 866 |
145 | 867 // Finally, subtract the multiplied divided value from the original |
868 result = new (phase->C, 3) SubINode( in(1), mult ); | |
869 } | |
0 | 870 |
871 // Now remove the bogus extra edges used to keep things alive | |
872 if (can_reshape) { | |
873 phase->is_IterGVN()->remove_dead_node(hook); | |
874 } else { | |
875 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
876 } | |
877 | |
878 // return the value | |
879 return result; | |
880 } | |
881 | |
882 //------------------------------Value------------------------------------------ | |
883 const Type *ModINode::Value( PhaseTransform *phase ) const { | |
884 // Either input is TOP ==> the result is TOP | |
885 const Type *t1 = phase->type( in(1) ); | |
886 const Type *t2 = phase->type( in(2) ); | |
887 if( t1 == Type::TOP ) return Type::TOP; | |
888 if( t2 == Type::TOP ) return Type::TOP; | |
889 | |
890 // We always generate the dynamic check for 0. | |
891 // 0 MOD X is 0 | |
892 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; | |
893 // X MOD X is 0 | |
894 if( phase->eqv( in(1), in(2) ) ) return TypeInt::ZERO; | |
895 | |
896 // Either input is BOTTOM ==> the result is the local BOTTOM | |
897 const Type *bot = bottom_type(); | |
898 if( (t1 == bot) || (t2 == bot) || | |
899 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
900 return bot; | |
901 | |
902 const TypeInt *i1 = t1->is_int(); | |
903 const TypeInt *i2 = t2->is_int(); | |
904 if( !i1->is_con() || !i2->is_con() ) { | |
905 if( i1->_lo >= 0 && i2->_lo >= 0 ) | |
906 return TypeInt::POS; | |
907 // If both numbers are not constants, we know little. | |
908 return TypeInt::INT; | |
909 } | |
910 // Mod by zero? Throw exception at runtime! | |
911 if( !i2->get_con() ) return TypeInt::POS; | |
912 | |
913 // We must be modulo'ing 2 float constants. | |
914 // Check for min_jint % '-1', result is defined to be '0'. | |
915 if( i1->get_con() == min_jint && i2->get_con() == -1 ) | |
916 return TypeInt::ZERO; | |
917 | |
918 return TypeInt::make( i1->get_con() % i2->get_con() ); | |
919 } | |
920 | |
921 | |
922 //============================================================================= | |
923 //------------------------------Idealize--------------------------------------- | |
924 Node *ModLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
925 // Check for dead control input | |
926 if( remove_dead_region(phase, can_reshape) ) return this; | |
927 | |
928 // Get the modulus | |
929 const Type *t = phase->type( in(2) ); | |
930 if( t == Type::TOP ) return NULL; | |
145 | 931 const TypeLong *tl = t->is_long(); |
0 | 932 |
933 // Check for useless control input | |
934 // Check for excluding mod-zero case | |
145 | 935 if( in(0) && (tl->_hi < 0 || tl->_lo > 0) ) { |
0 | 936 set_req(0, NULL); // Yank control input |
937 return this; | |
938 } | |
939 | |
940 // See if we are MOD'ing by 2^k or 2^k-1. | |
145 | 941 if( !tl->is_con() ) return NULL; |
942 jlong con = tl->get_con(); | |
943 | |
944 Node *hook = new (phase->C, 1) Node(1); | |
0 | 945 |
946 // Expand mod | |
145 | 947 if( con >= 0 && con < max_jlong && is_power_of_2_long(con+1) ) { |
948 uint k = log2_long(con); // Extract k | |
949 | |
0 | 950 // Basic algorithm by David Detlefs. See fastmod_long.java for gory details. |
951 // Used to help a popular random number generator which does a long-mod | |
952 // of 2^31-1 and shows up in SpecJBB and SciMark. | |
953 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*/}; | |
954 int trip_count = 1; | |
955 if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k]; | |
956 | |
145 | 957 // If the unroll factor is not too large, and if conditional moves are |
958 // ok, then use this case | |
959 if( trip_count <= 5 && ConditionalMoveLimit != 0 ) { | |
960 Node *x = in(1); // Value being mod'd | |
961 Node *divisor = in(2); // Also is mask | |
0 | 962 |
145 | 963 hook->init_req(0, x); // Add a use to x to prevent him from dying |
964 // Generate code to reduce X rapidly to nearly 2^k-1. | |
965 for( int i = 0; i < trip_count; i++ ) { | |
0 | 966 Node *xl = phase->transform( new (phase->C, 3) AndLNode(x,divisor) ); |
967 Node *xh = phase->transform( new (phase->C, 3) RShiftLNode(x,phase->intcon(k)) ); // Must be signed | |
968 x = phase->transform( new (phase->C, 3) AddLNode(xh,xl) ); | |
969 hook->set_req(0, x); // Add a use to x to prevent him from dying | |
145 | 970 } |
971 | |
972 // Generate sign-fixup code. Was original value positive? | |
973 // long hack_res = (i >= 0) ? divisor : CONST64(1); | |
974 Node *cmp1 = phase->transform( new (phase->C, 3) CmpLNode( in(1), phase->longcon(0) ) ); | |
975 Node *bol1 = phase->transform( new (phase->C, 2) BoolNode( cmp1, BoolTest::ge ) ); | |
976 Node *cmov1= phase->transform( new (phase->C, 4) CMoveLNode(bol1, phase->longcon(1), divisor, TypeLong::LONG) ); | |
977 // if( x >= hack_res ) x -= divisor; | |
978 Node *sub = phase->transform( new (phase->C, 3) SubLNode( x, divisor ) ); | |
979 Node *cmp2 = phase->transform( new (phase->C, 3) CmpLNode( x, cmov1 ) ); | |
980 Node *bol2 = phase->transform( new (phase->C, 2) BoolNode( cmp2, BoolTest::ge ) ); | |
981 // Convention is to not transform the return value of an Ideal | |
982 // since Ideal is expected to return a modified 'this' or a new node. | |
983 Node *cmov2= new (phase->C, 4) CMoveLNode(bol2, x, sub, TypeLong::LONG); | |
984 // cmov2 is now the mod | |
985 | |
986 // Now remove the bogus extra edges used to keep things alive | |
987 if (can_reshape) { | |
988 phase->is_IterGVN()->remove_dead_node(hook); | |
989 } else { | |
990 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
991 } | |
992 return cmov2; | |
0 | 993 } |
145 | 994 } |
995 | |
996 // Fell thru, the unroll case is not appropriate. Transform the modulo | |
997 // into a long multiply/int multiply/subtract case | |
998 | |
999 // Cannot handle mod 0, and min_jint isn't handled by the transform | |
1000 if( con == 0 || con == min_jlong ) return NULL; | |
1001 | |
1002 // Get the absolute value of the constant; at this point, we can use this | |
1003 jlong pos_con = (con >= 0) ? con : -con; | |
1004 | |
1005 // integer Mod 1 is always 0 | |
1006 if( pos_con == 1 ) return new (phase->C, 1) ConLNode(TypeLong::ZERO); | |
1007 | |
1008 int log2_con = -1; | |
1009 | |
1010 // If this is a power of two, they maybe we can mask it | |
1011 if( is_power_of_2_long(pos_con) ) { | |
1012 log2_con = log2_long(pos_con); | |
1013 | |
1014 const Type *dt = phase->type(in(1)); | |
1015 const TypeLong *dtl = dt->isa_long(); | |
1016 | |
1017 // See if this can be masked, if the dividend is non-negative | |
1018 if( dtl && dtl->_lo >= 0 ) | |
1019 return ( new (phase->C, 3) AndLNode( in(1), phase->longcon( pos_con-1 ) ) ); | |
1020 } | |
0 | 1021 |
145 | 1022 // Save in(1) so that it cannot be changed or deleted |
1023 hook->init_req(0, in(1)); | |
1024 | |
1025 // Divide using the transform from DivI to MulL | |
1026 Node *result = transform_long_divide( phase, in(1), pos_con ); | |
1027 if (result != NULL) { | |
1028 Node *divide = phase->transform(result); | |
1029 | |
1030 // Re-multiply, using a shift if this is a power of two | |
1031 Node *mult = NULL; | |
1032 | |
1033 if( log2_con >= 0 ) | |
1034 mult = phase->transform( new (phase->C, 3) LShiftLNode( divide, phase->intcon( log2_con ) ) ); | |
1035 else | |
1036 mult = phase->transform( new (phase->C, 3) MulLNode( divide, phase->longcon( pos_con ) ) ); | |
1037 | |
1038 // Finally, subtract the multiplied divided value from the original | |
1039 result = new (phase->C, 3) SubLNode( in(1), mult ); | |
0 | 1040 } |
145 | 1041 |
1042 // Now remove the bogus extra edges used to keep things alive | |
1043 if (can_reshape) { | |
1044 phase->is_IterGVN()->remove_dead_node(hook); | |
1045 } else { | |
1046 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
1047 } | |
1048 | |
1049 // return the value | |
1050 return result; | |
0 | 1051 } |
1052 | |
1053 //------------------------------Value------------------------------------------ | |
1054 const Type *ModLNode::Value( PhaseTransform *phase ) const { | |
1055 // Either input is TOP ==> the result is TOP | |
1056 const Type *t1 = phase->type( in(1) ); | |
1057 const Type *t2 = phase->type( in(2) ); | |
1058 if( t1 == Type::TOP ) return Type::TOP; | |
1059 if( t2 == Type::TOP ) return Type::TOP; | |
1060 | |
1061 // We always generate the dynamic check for 0. | |
1062 // 0 MOD X is 0 | |
1063 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; | |
1064 // X MOD X is 0 | |
1065 if( phase->eqv( in(1), in(2) ) ) return TypeLong::ZERO; | |
1066 | |
1067 // Either input is BOTTOM ==> the result is the local BOTTOM | |
1068 const Type *bot = bottom_type(); | |
1069 if( (t1 == bot) || (t2 == bot) || | |
1070 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
1071 return bot; | |
1072 | |
1073 const TypeLong *i1 = t1->is_long(); | |
1074 const TypeLong *i2 = t2->is_long(); | |
1075 if( !i1->is_con() || !i2->is_con() ) { | |
1076 if( i1->_lo >= CONST64(0) && i2->_lo >= CONST64(0) ) | |
1077 return TypeLong::POS; | |
1078 // If both numbers are not constants, we know little. | |
1079 return TypeLong::LONG; | |
1080 } | |
1081 // Mod by zero? Throw exception at runtime! | |
1082 if( !i2->get_con() ) return TypeLong::POS; | |
1083 | |
1084 // We must be modulo'ing 2 float constants. | |
1085 // Check for min_jint % '-1', result is defined to be '0'. | |
1086 if( i1->get_con() == min_jlong && i2->get_con() == -1 ) | |
1087 return TypeLong::ZERO; | |
1088 | |
1089 return TypeLong::make( i1->get_con() % i2->get_con() ); | |
1090 } | |
1091 | |
1092 | |
1093 //============================================================================= | |
1094 //------------------------------Value------------------------------------------ | |
1095 const Type *ModFNode::Value( PhaseTransform *phase ) const { | |
1096 // Either input is TOP ==> the result is TOP | |
1097 const Type *t1 = phase->type( in(1) ); | |
1098 const Type *t2 = phase->type( in(2) ); | |
1099 if( t1 == Type::TOP ) return Type::TOP; | |
1100 if( t2 == Type::TOP ) return Type::TOP; | |
1101 | |
1102 // Either input is BOTTOM ==> the result is the local BOTTOM | |
1103 const Type *bot = bottom_type(); | |
1104 if( (t1 == bot) || (t2 == bot) || | |
1105 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
1106 return bot; | |
1107 | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1108 // If either number is not a constant, we know nothing. |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1109 if ((t1->base() != Type::FloatCon) || (t2->base() != Type::FloatCon)) { |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1110 return Type::FLOAT; // note: x%x can be either NaN or 0 |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1111 } |
0 | 1112 |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1113 float f1 = t1->getf(); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1114 float f2 = t2->getf(); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1115 jint x1 = jint_cast(f1); // note: *(int*)&f1, not just (int)f1 |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1116 jint x2 = jint_cast(f2); |
0 | 1117 |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1118 // If either is a NaN, return an input NaN |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1119 if (g_isnan(f1)) return t1; |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1120 if (g_isnan(f2)) return t2; |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1121 |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1122 // If an operand is infinity or the divisor is +/- zero, punt. |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1123 if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jint) |
0 | 1124 return Type::FLOAT; |
1125 | |
1126 // We must be modulo'ing 2 float constants. | |
1127 // Make sure that the sign of the fmod is equal to the sign of the dividend | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1128 jint xr = jint_cast(fmod(f1, f2)); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1129 if ((x1 ^ xr) < 0) { |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1130 xr ^= min_jint; |
0 | 1131 } |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1132 |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1133 return TypeF::make(jfloat_cast(xr)); |
0 | 1134 } |
1135 | |
1136 | |
1137 //============================================================================= | |
1138 //------------------------------Value------------------------------------------ | |
1139 const Type *ModDNode::Value( PhaseTransform *phase ) const { | |
1140 // Either input is TOP ==> the result is TOP | |
1141 const Type *t1 = phase->type( in(1) ); | |
1142 const Type *t2 = phase->type( in(2) ); | |
1143 if( t1 == Type::TOP ) return Type::TOP; | |
1144 if( t2 == Type::TOP ) return Type::TOP; | |
1145 | |
1146 // Either input is BOTTOM ==> the result is the local BOTTOM | |
1147 const Type *bot = bottom_type(); | |
1148 if( (t1 == bot) || (t2 == bot) || | |
1149 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
1150 return bot; | |
1151 | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1152 // If either number is not a constant, we know nothing. |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1153 if ((t1->base() != Type::DoubleCon) || (t2->base() != Type::DoubleCon)) { |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1154 return Type::DOUBLE; // note: x%x can be either NaN or 0 |
0 | 1155 } |
1156 | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1157 double f1 = t1->getd(); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1158 double f2 = t2->getd(); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1159 jlong x1 = jlong_cast(f1); // note: *(long*)&f1, not just (long)f1 |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1160 jlong x2 = jlong_cast(f2); |
0 | 1161 |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1162 // If either is a NaN, return an input NaN |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1163 if (g_isnan(f1)) return t1; |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1164 if (g_isnan(f2)) return t2; |
0 | 1165 |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1166 // If an operand is infinity or the divisor is +/- zero, punt. |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1167 if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jlong) |
0 | 1168 return Type::DOUBLE; |
1169 | |
1170 // We must be modulo'ing 2 double constants. | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1171 // Make sure that the sign of the fmod is equal to the sign of the dividend |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1172 jlong xr = jlong_cast(fmod(f1, f2)); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1173 if ((x1 ^ xr) < 0) { |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1174 xr ^= min_jlong; |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1175 } |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1176 |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1177 return TypeD::make(jdouble_cast(xr)); |
0 | 1178 } |
1179 | |
1180 //============================================================================= | |
1181 | |
1182 DivModNode::DivModNode( Node *c, Node *dividend, Node *divisor ) : MultiNode(3) { | |
1183 init_req(0, c); | |
1184 init_req(1, dividend); | |
1185 init_req(2, divisor); | |
1186 } | |
1187 | |
1188 //------------------------------make------------------------------------------ | |
1189 DivModINode* DivModINode::make(Compile* C, Node* div_or_mod) { | |
1190 Node* n = div_or_mod; | |
1191 assert(n->Opcode() == Op_DivI || n->Opcode() == Op_ModI, | |
1192 "only div or mod input pattern accepted"); | |
1193 | |
1194 DivModINode* divmod = new (C, 3) DivModINode(n->in(0), n->in(1), n->in(2)); | |
1195 Node* dproj = new (C, 1) ProjNode(divmod, DivModNode::div_proj_num); | |
1196 Node* mproj = new (C, 1) ProjNode(divmod, DivModNode::mod_proj_num); | |
1197 return divmod; | |
1198 } | |
1199 | |
1200 //------------------------------make------------------------------------------ | |
1201 DivModLNode* DivModLNode::make(Compile* C, Node* div_or_mod) { | |
1202 Node* n = div_or_mod; | |
1203 assert(n->Opcode() == Op_DivL || n->Opcode() == Op_ModL, | |
1204 "only div or mod input pattern accepted"); | |
1205 | |
1206 DivModLNode* divmod = new (C, 3) DivModLNode(n->in(0), n->in(1), n->in(2)); | |
1207 Node* dproj = new (C, 1) ProjNode(divmod, DivModNode::div_proj_num); | |
1208 Node* mproj = new (C, 1) ProjNode(divmod, DivModNode::mod_proj_num); | |
1209 return divmod; | |
1210 } | |
1211 | |
1212 //------------------------------match------------------------------------------ | |
1213 // return result(s) along with their RegMask info | |
1214 Node *DivModINode::match( const ProjNode *proj, const Matcher *match ) { | |
1215 uint ideal_reg = proj->ideal_reg(); | |
1216 RegMask rm; | |
1217 if (proj->_con == div_proj_num) { | |
1218 rm = match->divI_proj_mask(); | |
1219 } else { | |
1220 assert(proj->_con == mod_proj_num, "must be div or mod projection"); | |
1221 rm = match->modI_proj_mask(); | |
1222 } | |
1223 return new (match->C, 1)MachProjNode(this, proj->_con, rm, ideal_reg); | |
1224 } | |
1225 | |
1226 | |
1227 //------------------------------match------------------------------------------ | |
1228 // return result(s) along with their RegMask info | |
1229 Node *DivModLNode::match( const ProjNode *proj, const Matcher *match ) { | |
1230 uint ideal_reg = proj->ideal_reg(); | |
1231 RegMask rm; | |
1232 if (proj->_con == div_proj_num) { | |
1233 rm = match->divL_proj_mask(); | |
1234 } else { | |
1235 assert(proj->_con == mod_proj_num, "must be div or mod projection"); | |
1236 rm = match->modL_proj_mask(); | |
1237 } | |
1238 return new (match->C, 1)MachProjNode(this, proj->_con, rm, ideal_reg); | |
1239 } |