Mercurial > hg > graal-compiler
annotate src/share/vm/opto/divnode.cpp @ 775:93c14e5562c4
6823354: Add intrinsics for {Integer,Long}.{numberOfLeadingZeros,numberOfTrailingZeros}()
Summary: These methods can be instrinsified by using bit scan, bit test, and population count instructions.
Reviewed-by: kvn, never
author | twisti |
---|---|
date | Wed, 06 May 2009 00:27:52 -0700 |
parents | 98cb887364d3 |
children | cecd04fc6f93 |
rev | line source |
---|---|
0 | 1 /* |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
2 * Copyright 1997-2009 Sun Microsystems, Inc. All Rights Reserved. |
0 | 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 // | |
605 | 38 // Borrowed almost verbatim from Hacker's Delight by Henry S. Warren, Jr. with |
145 | 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. | |
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
113 const TypeInt *andconi_t = phase->type( dividend->in(2) )->isa_int(); |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
114 if( andconi_t && andconi_t->is_con() ) { |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
115 jint andconi = andconi_t->get_con(); |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
116 if( andconi < 0 && is_power_of_2(-andconi) && (-andconi) >= d ) { |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
117 dividend = dividend->in(1); |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
118 needs_rounding = false; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
119 } |
0 | 120 } |
121 } | |
122 | |
123 // Add rounding to the shift to handle the sign bit | |
145 | 124 int l = log2_intptr(d-1)+1; |
125 if (needs_rounding) { | |
126 // Divide-by-power-of-2 can be made into a shift, but you have to do | |
127 // more math for the rounding. You need to add 0 for positive | |
128 // numbers, and "i-1" for negative numbers. Example: i=4, so the | |
129 // shift is by 2. You need to add 3 to negative dividends and 0 to | |
130 // positive ones. So (-7+3)>>2 becomes -1, (-4+3)>>2 becomes -1, | |
131 // (-2+3)>>2 becomes 0, etc. | |
132 | |
133 // Compute 0 or -1, based on sign bit | |
134 Node *sign = phase->transform(new (phase->C, 3) RShiftINode(dividend, phase->intcon(N - 1))); | |
135 // Mask sign bit to the low sign bits | |
136 Node *round = phase->transform(new (phase->C, 3) URShiftINode(sign, phase->intcon(N - l))); | |
137 // Round up before shifting | |
138 dividend = phase->transform(new (phase->C, 3) AddINode(dividend, round)); | |
0 | 139 } |
140 | |
145 | 141 // Shift for division |
0 | 142 q = new (phase->C, 3) RShiftINode(dividend, phase->intcon(l)); |
143 | |
145 | 144 if (!d_pos) { |
0 | 145 q = new (phase->C, 3) SubINode(phase->intcon(0), phase->transform(q)); |
145 | 146 } |
147 } else { | |
148 // Attempt the jint constant divide -> multiply transform found in | |
149 // "Division by Invariant Integers using Multiplication" | |
150 // by Granlund and Montgomery | |
151 // See also "Hacker's Delight", chapter 10 by Warren. | |
152 | |
153 jint magic_const; | |
154 jint shift_const; | |
155 if (magic_int_divide_constants(d, magic_const, shift_const)) { | |
156 Node *magic = phase->longcon(magic_const); | |
157 Node *dividend_long = phase->transform(new (phase->C, 2) ConvI2LNode(dividend)); | |
158 | |
159 // Compute the high half of the dividend x magic multiplication | |
160 Node *mul_hi = phase->transform(new (phase->C, 3) MulLNode(dividend_long, magic)); | |
161 | |
162 if (magic_const < 0) { | |
163 mul_hi = phase->transform(new (phase->C, 3) RShiftLNode(mul_hi, phase->intcon(N))); | |
164 mul_hi = phase->transform(new (phase->C, 2) ConvL2INode(mul_hi)); | |
165 | |
166 // The magic multiplier is too large for a 32 bit constant. We've adjusted | |
167 // it down by 2^32, but have to add 1 dividend back in after the multiplication. | |
168 // This handles the "overflow" case described by Granlund and Montgomery. | |
169 mul_hi = phase->transform(new (phase->C, 3) AddINode(dividend, mul_hi)); | |
170 | |
171 // Shift over the (adjusted) mulhi | |
172 if (shift_const != 0) { | |
173 mul_hi = phase->transform(new (phase->C, 3) RShiftINode(mul_hi, phase->intcon(shift_const))); | |
174 } | |
175 } else { | |
176 // No add is required, we can merge the shifts together. | |
177 mul_hi = phase->transform(new (phase->C, 3) RShiftLNode(mul_hi, phase->intcon(N + shift_const))); | |
178 mul_hi = phase->transform(new (phase->C, 2) ConvL2INode(mul_hi)); | |
179 } | |
180 | |
181 // Get a 0 or -1 from the sign of the dividend. | |
182 Node *addend0 = mul_hi; | |
183 Node *addend1 = phase->transform(new (phase->C, 3) RShiftINode(dividend, phase->intcon(N-1))); | |
184 | |
185 // If the divisor is negative, swap the order of the input addends; | |
186 // this has the effect of negating the quotient. | |
187 if (!d_pos) { | |
188 Node *temp = addend0; addend0 = addend1; addend1 = temp; | |
189 } | |
190 | |
191 // Adjust the final quotient by subtracting -1 (adding 1) | |
192 // from the mul_hi. | |
193 q = new (phase->C, 3) SubINode(addend0, addend1); | |
194 } | |
195 } | |
196 | |
197 return q; | |
198 } | |
199 | |
200 //---------------------magic_long_divide_constants----------------------------- | |
201 // Compute magic multiplier and shift constant for converting a 64 bit divide | |
202 // by constant into a multiply/shift/add series. Return false if calculations | |
203 // fail. | |
204 // | |
605 | 205 // Borrowed almost verbatim from Hacker's Delight by Henry S. Warren, Jr. with |
145 | 206 // minor type name and parameter changes. Adjusted to 64 bit word width. |
207 static bool magic_long_divide_constants(jlong d, jlong &M, jint &s) { | |
208 int64_t p; | |
209 uint64_t ad, anc, delta, q1, r1, q2, r2, t; | |
210 const uint64_t two63 = 0x8000000000000000LL; // 2**63. | |
211 | |
212 ad = ABS(d); | |
213 if (d == 0 || d == 1) return false; | |
214 t = two63 + ((uint64_t)d >> 63); | |
215 anc = t - 1 - t%ad; // Absolute value of nc. | |
216 p = 63; // Init. p. | |
217 q1 = two63/anc; // Init. q1 = 2**p/|nc|. | |
218 r1 = two63 - q1*anc; // Init. r1 = rem(2**p, |nc|). | |
219 q2 = two63/ad; // Init. q2 = 2**p/|d|. | |
220 r2 = two63 - q2*ad; // Init. r2 = rem(2**p, |d|). | |
221 do { | |
222 p = p + 1; | |
223 q1 = 2*q1; // Update q1 = 2**p/|nc|. | |
224 r1 = 2*r1; // Update r1 = rem(2**p, |nc|). | |
225 if (r1 >= anc) { // (Must be an unsigned | |
226 q1 = q1 + 1; // comparison here). | |
227 r1 = r1 - anc; | |
228 } | |
229 q2 = 2*q2; // Update q2 = 2**p/|d|. | |
230 r2 = 2*r2; // Update r2 = rem(2**p, |d|). | |
231 if (r2 >= ad) { // (Must be an unsigned | |
232 q2 = q2 + 1; // comparison here). | |
233 r2 = r2 - ad; | |
234 } | |
235 delta = ad - r2; | |
236 } while (q1 < delta || (q1 == delta && r1 == 0)); | |
237 | |
238 M = q2 + 1; | |
239 if (d < 0) M = -M; // Magic number and | |
240 s = p - 64; // shift amount to return. | |
241 | |
242 return true; | |
243 } | |
244 | |
245 //---------------------long_by_long_mulhi-------------------------------------- | |
246 // Generate ideal node graph for upper half of a 64 bit x 64 bit multiplication | |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
247 static Node* long_by_long_mulhi(PhaseGVN* phase, Node* dividend, jlong magic_const) { |
145 | 248 // If the architecture supports a 64x64 mulhi, there is |
249 // no need to synthesize it in ideal nodes. | |
250 if (Matcher::has_match_rule(Op_MulHiL)) { | |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
251 Node* v = phase->longcon(magic_const); |
145 | 252 return new (phase->C, 3) MulHiLNode(dividend, v); |
0 | 253 } |
254 | |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
255 // Taken from Hacker's Delight, Fig. 8-2. Multiply high signed. |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
256 // (http://www.hackersdelight.org/HDcode/mulhs.c) |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
257 // |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
258 // int mulhs(int u, int v) { |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
259 // unsigned u0, v0, w0; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
260 // int u1, v1, w1, w2, t; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
261 // |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
262 // u0 = u & 0xFFFF; u1 = u >> 16; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
263 // v0 = v & 0xFFFF; v1 = v >> 16; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
264 // w0 = u0*v0; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
265 // t = u1*v0 + (w0 >> 16); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
266 // w1 = t & 0xFFFF; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
267 // w2 = t >> 16; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
268 // w1 = u0*v1 + w1; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
269 // return u1*v1 + w2 + (w1 >> 16); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
270 // } |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
271 // |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
272 // Note: The version above is for 32x32 multiplications, while the |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
273 // following inline comments are adapted to 64x64. |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
274 |
145 | 275 const int N = 64; |
276 | |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
277 // u0 = u & 0xFFFFFFFF; u1 = u >> 32; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
278 Node* u0 = phase->transform(new (phase->C, 3) AndLNode(dividend, phase->longcon(0xFFFFFFFF))); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
279 Node* u1 = phase->transform(new (phase->C, 3) RShiftLNode(dividend, phase->intcon(N / 2))); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
280 |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
281 // v0 = v & 0xFFFFFFFF; v1 = v >> 32; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
282 Node* v0 = phase->longcon(magic_const & 0xFFFFFFFF); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
283 Node* v1 = phase->longcon(magic_const >> (N / 2)); |
145 | 284 |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
285 // w0 = u0*v0; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
286 Node* w0 = phase->transform(new (phase->C, 3) MulLNode(u0, v0)); |
145 | 287 |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
288 // t = u1*v0 + (w0 >> 32); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
289 Node* u1v0 = phase->transform(new (phase->C, 3) MulLNode(u1, v0)); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
290 Node* temp = phase->transform(new (phase->C, 3) URShiftLNode(w0, phase->intcon(N / 2))); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
291 Node* t = phase->transform(new (phase->C, 3) AddLNode(u1v0, temp)); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
292 |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
293 // w1 = t & 0xFFFFFFFF; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
294 Node* w1 = new (phase->C, 3) AndLNode(t, phase->longcon(0xFFFFFFFF)); |
145 | 295 |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
296 // w2 = t >> 32; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
297 Node* w2 = new (phase->C, 3) RShiftLNode(t, phase->intcon(N / 2)); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
298 |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
299 // 6732154: Construct both w1 and w2 before transforming, so t |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
300 // doesn't go dead prematurely. |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
301 w1 = phase->transform(w1); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
302 w2 = phase->transform(w2); |
294
616a07a75c3c
6732154: REG: Printing an Image using image/gif doc flavor crashes the VM, Solsparc
rasbold
parents:
196
diff
changeset
|
303 |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
304 // w1 = u0*v1 + w1; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
305 Node* u0v1 = phase->transform(new (phase->C, 3) MulLNode(u0, v1)); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
306 w1 = phase->transform(new (phase->C, 3) AddLNode(u0v1, w1)); |
294
616a07a75c3c
6732154: REG: Printing an Image using image/gif doc flavor crashes the VM, Solsparc
rasbold
parents:
196
diff
changeset
|
307 |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
308 // return u1*v1 + w2 + (w1 >> 32); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
309 Node* u1v1 = phase->transform(new (phase->C, 3) MulLNode(u1, v1)); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
310 Node* temp1 = phase->transform(new (phase->C, 3) AddLNode(u1v1, w2)); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
311 Node* temp2 = phase->transform(new (phase->C, 3) RShiftLNode(w1, phase->intcon(N / 2))); |
145 | 312 |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
313 return new (phase->C, 3) AddLNode(temp1, temp2); |
145 | 314 } |
315 | |
316 | |
317 //--------------------------transform_long_divide------------------------------ | |
318 // Convert a division by constant divisor into an alternate Ideal graph. | |
319 // Return NULL if no transformation occurs. | |
320 static Node *transform_long_divide( PhaseGVN *phase, Node *dividend, jlong divisor ) { | |
321 // Check for invalid divisors | |
322 assert( divisor != 0L && divisor != min_jlong, | |
323 "bad divisor for transforming to long multiply" ); | |
324 | |
325 bool d_pos = divisor >= 0; | |
326 jlong d = d_pos ? divisor : -divisor; | |
327 const int N = 64; | |
328 | |
329 // Result | |
330 Node *q = NULL; | |
331 | |
332 if (d == 1) { | |
333 // division by +/- 1 | |
334 if (!d_pos) { | |
335 // Just negate the value | |
336 q = new (phase->C, 3) SubLNode(phase->longcon(0), dividend); | |
337 } | |
338 } else if ( is_power_of_2_long(d) ) { | |
339 | |
340 // division by +/- a power of 2 | |
341 | |
342 // See if we can simply do a shift without rounding | |
343 bool needs_rounding = true; | |
344 const Type *dt = phase->type(dividend); | |
345 const TypeLong *dtl = dt->isa_long(); | |
0 | 346 |
145 | 347 if (dtl && dtl->_lo > 0) { |
348 // we don't need to round a positive dividend | |
349 needs_rounding = false; | |
350 } else if( dividend->Opcode() == Op_AndL ) { | |
351 // An AND mask of sufficient size clears the low bits and | |
352 // I can avoid rounding. | |
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
353 const TypeLong *andconl_t = phase->type( dividend->in(2) )->isa_long(); |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
354 if( andconl_t && andconl_t->is_con() ) { |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
355 jlong andconl = andconl_t->get_con(); |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
356 if( andconl < 0 && is_power_of_2_long(-andconl) && (-andconl) >= d ) { |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
357 dividend = dividend->in(1); |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
358 needs_rounding = false; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
359 } |
145 | 360 } |
361 } | |
362 | |
363 // Add rounding to the shift to handle the sign bit | |
364 int l = log2_long(d-1)+1; | |
365 if (needs_rounding) { | |
366 // Divide-by-power-of-2 can be made into a shift, but you have to do | |
367 // more math for the rounding. You need to add 0 for positive | |
368 // numbers, and "i-1" for negative numbers. Example: i=4, so the | |
369 // shift is by 2. You need to add 3 to negative dividends and 0 to | |
370 // positive ones. So (-7+3)>>2 becomes -1, (-4+3)>>2 becomes -1, | |
371 // (-2+3)>>2 becomes 0, etc. | |
372 | |
373 // Compute 0 or -1, based on sign bit | |
374 Node *sign = phase->transform(new (phase->C, 3) RShiftLNode(dividend, phase->intcon(N - 1))); | |
375 // Mask sign bit to the low sign bits | |
376 Node *round = phase->transform(new (phase->C, 3) URShiftLNode(sign, phase->intcon(N - l))); | |
377 // Round up before shifting | |
378 dividend = phase->transform(new (phase->C, 3) AddLNode(dividend, round)); | |
379 } | |
380 | |
381 // Shift for division | |
382 q = new (phase->C, 3) RShiftLNode(dividend, phase->intcon(l)); | |
383 | |
384 if (!d_pos) { | |
385 q = new (phase->C, 3) SubLNode(phase->longcon(0), phase->transform(q)); | |
386 } | |
387 } else { | |
388 // Attempt the jlong constant divide -> multiply transform found in | |
389 // "Division by Invariant Integers using Multiplication" | |
390 // by Granlund and Montgomery | |
391 // See also "Hacker's Delight", chapter 10 by Warren. | |
392 | |
393 jlong magic_const; | |
394 jint shift_const; | |
395 if (magic_long_divide_constants(d, magic_const, shift_const)) { | |
396 // Compute the high half of the dividend x magic multiplication | |
397 Node *mul_hi = phase->transform(long_by_long_mulhi(phase, dividend, magic_const)); | |
398 | |
399 // The high half of the 128-bit multiply is computed. | |
400 if (magic_const < 0) { | |
401 // The magic multiplier is too large for a 64 bit constant. We've adjusted | |
402 // it down by 2^64, but have to add 1 dividend back in after the multiplication. | |
403 // This handles the "overflow" case described by Granlund and Montgomery. | |
404 mul_hi = phase->transform(new (phase->C, 3) AddLNode(dividend, mul_hi)); | |
405 } | |
406 | |
407 // Shift over the (adjusted) mulhi | |
408 if (shift_const != 0) { | |
409 mul_hi = phase->transform(new (phase->C, 3) RShiftLNode(mul_hi, phase->intcon(shift_const))); | |
410 } | |
411 | |
412 // Get a 0 or -1 from the sign of the dividend. | |
413 Node *addend0 = mul_hi; | |
414 Node *addend1 = phase->transform(new (phase->C, 3) RShiftLNode(dividend, phase->intcon(N-1))); | |
415 | |
416 // If the divisor is negative, swap the order of the input addends; | |
417 // this has the effect of negating the quotient. | |
418 if (!d_pos) { | |
419 Node *temp = addend0; addend0 = addend1; addend1 = temp; | |
420 } | |
421 | |
422 // Adjust the final quotient by subtracting -1 (adding 1) | |
423 // from the mul_hi. | |
424 q = new (phase->C, 3) SubLNode(addend0, addend1); | |
425 } | |
0 | 426 } |
427 | |
145 | 428 return q; |
0 | 429 } |
430 | |
431 //============================================================================= | |
432 //------------------------------Identity--------------------------------------- | |
433 // If the divisor is 1, we are an identity on the dividend. | |
434 Node *DivINode::Identity( PhaseTransform *phase ) { | |
435 return (phase->type( in(2) )->higher_equal(TypeInt::ONE)) ? in(1) : this; | |
436 } | |
437 | |
438 //------------------------------Idealize--------------------------------------- | |
439 // Divides can be changed to multiplies and/or shifts | |
440 Node *DivINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
441 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
305 | 442 // Don't bother trying to transform a dead node |
443 if( in(0) && in(0)->is_top() ) return NULL; | |
0 | 444 |
445 const Type *t = phase->type( in(2) ); | |
446 if( t == TypeInt::ONE ) // Identity? | |
447 return NULL; // Skip it | |
448 | |
449 const TypeInt *ti = t->isa_int(); | |
450 if( !ti ) return NULL; | |
451 if( !ti->is_con() ) return NULL; | |
145 | 452 jint i = ti->get_con(); // Get divisor |
0 | 453 |
454 if (i == 0) return NULL; // Dividing by zero constant does not idealize | |
455 | |
456 set_req(0,NULL); // Dividing by a not-zero constant; no faulting | |
457 | |
458 // Dividing by MININT does not optimize as a power-of-2 shift. | |
459 if( i == min_jint ) return NULL; | |
460 | |
145 | 461 return transform_int_divide( phase, in(1), i ); |
0 | 462 } |
463 | |
464 //------------------------------Value------------------------------------------ | |
465 // A DivINode divides its inputs. The third input is a Control input, used to | |
466 // prevent hoisting the divide above an unsafe test. | |
467 const Type *DivINode::Value( PhaseTransform *phase ) const { | |
468 // Either input is TOP ==> the result is TOP | |
469 const Type *t1 = phase->type( in(1) ); | |
470 const Type *t2 = phase->type( in(2) ); | |
471 if( t1 == Type::TOP ) return Type::TOP; | |
472 if( t2 == Type::TOP ) return Type::TOP; | |
473 | |
474 // x/x == 1 since we always generate the dynamic divisor check for 0. | |
475 if( phase->eqv( in(1), in(2) ) ) | |
476 return TypeInt::ONE; | |
477 | |
478 // Either input is BOTTOM ==> the result is the local BOTTOM | |
479 const Type *bot = bottom_type(); | |
480 if( (t1 == bot) || (t2 == bot) || | |
481 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
482 return bot; | |
483 | |
484 // Divide the two numbers. We approximate. | |
485 // If divisor is a constant and not zero | |
486 const TypeInt *i1 = t1->is_int(); | |
487 const TypeInt *i2 = t2->is_int(); | |
488 int widen = MAX2(i1->_widen, i2->_widen); | |
489 | |
490 if( i2->is_con() && i2->get_con() != 0 ) { | |
491 int32 d = i2->get_con(); // Divisor | |
492 jint lo, hi; | |
493 if( d >= 0 ) { | |
494 lo = i1->_lo/d; | |
495 hi = i1->_hi/d; | |
496 } else { | |
497 if( d == -1 && i1->_lo == min_jint ) { | |
498 // 'min_jint/-1' throws arithmetic exception during compilation | |
499 lo = min_jint; | |
500 // do not support holes, 'hi' must go to either min_jint or max_jint: | |
501 // [min_jint, -10]/[-1,-1] ==> [min_jint] UNION [10,max_jint] | |
502 hi = i1->_hi == min_jint ? min_jint : max_jint; | |
503 } else { | |
504 lo = i1->_hi/d; | |
505 hi = i1->_lo/d; | |
506 } | |
507 } | |
508 return TypeInt::make(lo, hi, widen); | |
509 } | |
510 | |
511 // If the dividend is a constant | |
512 if( i1->is_con() ) { | |
513 int32 d = i1->get_con(); | |
514 if( d < 0 ) { | |
515 if( d == min_jint ) { | |
516 // (-min_jint) == min_jint == (min_jint / -1) | |
517 return TypeInt::make(min_jint, max_jint/2 + 1, widen); | |
518 } else { | |
519 return TypeInt::make(d, -d, widen); | |
520 } | |
521 } | |
522 return TypeInt::make(-d, d, widen); | |
523 } | |
524 | |
525 // Otherwise we give up all hope | |
526 return TypeInt::INT; | |
527 } | |
528 | |
529 | |
530 //============================================================================= | |
531 //------------------------------Identity--------------------------------------- | |
532 // If the divisor is 1, we are an identity on the dividend. | |
533 Node *DivLNode::Identity( PhaseTransform *phase ) { | |
534 return (phase->type( in(2) )->higher_equal(TypeLong::ONE)) ? in(1) : this; | |
535 } | |
536 | |
537 //------------------------------Idealize--------------------------------------- | |
538 // Dividing by a power of 2 is a shift. | |
539 Node *DivLNode::Ideal( PhaseGVN *phase, bool can_reshape) { | |
540 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
305 | 541 // Don't bother trying to transform a dead node |
542 if( in(0) && in(0)->is_top() ) return NULL; | |
0 | 543 |
544 const Type *t = phase->type( in(2) ); | |
145 | 545 if( t == TypeLong::ONE ) // Identity? |
0 | 546 return NULL; // Skip it |
547 | |
145 | 548 const TypeLong *tl = t->isa_long(); |
549 if( !tl ) return NULL; | |
550 if( !tl->is_con() ) return NULL; | |
551 jlong l = tl->get_con(); // Get divisor | |
552 | |
553 if (l == 0) return NULL; // Dividing by zero constant does not idealize | |
554 | |
555 set_req(0,NULL); // Dividing by a not-zero constant; no faulting | |
0 | 556 |
557 // Dividing by MININT does not optimize as a power-of-2 shift. | |
145 | 558 if( l == min_jlong ) return NULL; |
0 | 559 |
145 | 560 return transform_long_divide( phase, in(1), l ); |
0 | 561 } |
562 | |
563 //------------------------------Value------------------------------------------ | |
564 // A DivLNode divides its inputs. The third input is a Control input, used to | |
565 // prevent hoisting the divide above an unsafe test. | |
566 const Type *DivLNode::Value( PhaseTransform *phase ) const { | |
567 // Either input is TOP ==> the result is TOP | |
568 const Type *t1 = phase->type( in(1) ); | |
569 const Type *t2 = phase->type( in(2) ); | |
570 if( t1 == Type::TOP ) return Type::TOP; | |
571 if( t2 == Type::TOP ) return Type::TOP; | |
572 | |
573 // x/x == 1 since we always generate the dynamic divisor check for 0. | |
574 if( phase->eqv( in(1), in(2) ) ) | |
575 return TypeLong::ONE; | |
576 | |
577 // Either input is BOTTOM ==> the result is the local BOTTOM | |
578 const Type *bot = bottom_type(); | |
579 if( (t1 == bot) || (t2 == bot) || | |
580 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
581 return bot; | |
582 | |
583 // Divide the two numbers. We approximate. | |
584 // If divisor is a constant and not zero | |
585 const TypeLong *i1 = t1->is_long(); | |
586 const TypeLong *i2 = t2->is_long(); | |
587 int widen = MAX2(i1->_widen, i2->_widen); | |
588 | |
589 if( i2->is_con() && i2->get_con() != 0 ) { | |
590 jlong d = i2->get_con(); // Divisor | |
591 jlong lo, hi; | |
592 if( d >= 0 ) { | |
593 lo = i1->_lo/d; | |
594 hi = i1->_hi/d; | |
595 } else { | |
596 if( d == CONST64(-1) && i1->_lo == min_jlong ) { | |
597 // 'min_jlong/-1' throws arithmetic exception during compilation | |
598 lo = min_jlong; | |
599 // do not support holes, 'hi' must go to either min_jlong or max_jlong: | |
600 // [min_jlong, -10]/[-1,-1] ==> [min_jlong] UNION [10,max_jlong] | |
601 hi = i1->_hi == min_jlong ? min_jlong : max_jlong; | |
602 } else { | |
603 lo = i1->_hi/d; | |
604 hi = i1->_lo/d; | |
605 } | |
606 } | |
607 return TypeLong::make(lo, hi, widen); | |
608 } | |
609 | |
610 // If the dividend is a constant | |
611 if( i1->is_con() ) { | |
612 jlong d = i1->get_con(); | |
613 if( d < 0 ) { | |
614 if( d == min_jlong ) { | |
615 // (-min_jlong) == min_jlong == (min_jlong / -1) | |
616 return TypeLong::make(min_jlong, max_jlong/2 + 1, widen); | |
617 } else { | |
618 return TypeLong::make(d, -d, widen); | |
619 } | |
620 } | |
621 return TypeLong::make(-d, d, widen); | |
622 } | |
623 | |
624 // Otherwise we give up all hope | |
625 return TypeLong::LONG; | |
626 } | |
627 | |
628 | |
629 //============================================================================= | |
630 //------------------------------Value------------------------------------------ | |
631 // An DivFNode divides its inputs. The third input is a Control input, used to | |
632 // prevent hoisting the divide above an unsafe test. | |
633 const Type *DivFNode::Value( PhaseTransform *phase ) const { | |
634 // Either input is TOP ==> the result is TOP | |
635 const Type *t1 = phase->type( in(1) ); | |
636 const Type *t2 = phase->type( in(2) ); | |
637 if( t1 == Type::TOP ) return Type::TOP; | |
638 if( t2 == Type::TOP ) return Type::TOP; | |
639 | |
640 // Either input is BOTTOM ==> the result is the local BOTTOM | |
641 const Type *bot = bottom_type(); | |
642 if( (t1 == bot) || (t2 == bot) || | |
643 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
644 return bot; | |
645 | |
646 // x/x == 1, we ignore 0/0. | |
647 // 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
|
648 // Does not work for variables because of NaN's |
0 | 649 if( phase->eqv( in(1), in(2) ) && t1->base() == Type::FloatCon) |
650 if (!g_isnan(t1->getf()) && g_isfinite(t1->getf()) && t1->getf() != 0.0) // could be negative ZERO or NaN | |
651 return TypeF::ONE; | |
652 | |
653 if( t2 == TypeF::ONE ) | |
654 return t1; | |
655 | |
656 // If divisor is a constant and not zero, divide them numbers | |
657 if( t1->base() == Type::FloatCon && | |
658 t2->base() == Type::FloatCon && | |
659 t2->getf() != 0.0 ) // could be negative zero | |
660 return TypeF::make( t1->getf()/t2->getf() ); | |
661 | |
662 // If the dividend is a constant zero | |
663 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
664 // Test TypeF::ZERO is not sufficient as it could be negative zero | |
665 | |
666 if( t1 == TypeF::ZERO && !g_isnan(t2->getf()) && t2->getf() != 0.0 ) | |
667 return TypeF::ZERO; | |
668 | |
669 // Otherwise we give up all hope | |
670 return Type::FLOAT; | |
671 } | |
672 | |
673 //------------------------------isA_Copy--------------------------------------- | |
674 // Dividing by self is 1. | |
675 // If the divisor is 1, we are an identity on the dividend. | |
676 Node *DivFNode::Identity( PhaseTransform *phase ) { | |
677 return (phase->type( in(2) ) == TypeF::ONE) ? in(1) : this; | |
678 } | |
679 | |
680 | |
681 //------------------------------Idealize--------------------------------------- | |
682 Node *DivFNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
683 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
305 | 684 // Don't bother trying to transform a dead node |
685 if( in(0) && in(0)->is_top() ) return NULL; | |
0 | 686 |
687 const Type *t2 = phase->type( in(2) ); | |
688 if( t2 == TypeF::ONE ) // Identity? | |
689 return NULL; // Skip it | |
690 | |
691 const TypeF *tf = t2->isa_float_constant(); | |
692 if( !tf ) return NULL; | |
693 if( tf->base() != Type::FloatCon ) return NULL; | |
694 | |
695 // Check for out of range values | |
696 if( tf->is_nan() || !tf->is_finite() ) return NULL; | |
697 | |
698 // Get the value | |
699 float f = tf->getf(); | |
700 int exp; | |
701 | |
702 // Only for special case of dividing by a power of 2 | |
703 if( frexp((double)f, &exp) != 0.5 ) return NULL; | |
704 | |
705 // Limit the range of acceptable exponents | |
706 if( exp < -126 || exp > 126 ) return NULL; | |
707 | |
708 // Compute the reciprocal | |
709 float reciprocal = ((float)1.0) / f; | |
710 | |
711 assert( frexp((double)reciprocal, &exp) == 0.5, "reciprocal should be power of 2" ); | |
712 | |
713 // return multiplication by the reciprocal | |
714 return (new (phase->C, 3) MulFNode(in(1), phase->makecon(TypeF::make(reciprocal)))); | |
715 } | |
716 | |
717 //============================================================================= | |
718 //------------------------------Value------------------------------------------ | |
719 // 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
|
720 // prevent hoisting the divide above an unsafe test. |
0 | 721 const Type *DivDNode::Value( PhaseTransform *phase ) const { |
722 // Either input is TOP ==> the result is TOP | |
723 const Type *t1 = phase->type( in(1) ); | |
724 const Type *t2 = phase->type( in(2) ); | |
725 if( t1 == Type::TOP ) return Type::TOP; | |
726 if( t2 == Type::TOP ) return Type::TOP; | |
727 | |
728 // Either input is BOTTOM ==> the result is the local BOTTOM | |
729 const Type *bot = bottom_type(); | |
730 if( (t1 == bot) || (t2 == bot) || | |
731 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
732 return bot; | |
733 | |
734 // x/x == 1, we ignore 0/0. | |
735 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
736 // Does not work for variables because of NaN's | |
737 if( phase->eqv( in(1), in(2) ) && t1->base() == Type::DoubleCon) | |
738 if (!g_isnan(t1->getd()) && g_isfinite(t1->getd()) && t1->getd() != 0.0) // could be negative ZERO or NaN | |
739 return TypeD::ONE; | |
740 | |
741 if( t2 == TypeD::ONE ) | |
742 return t1; | |
743 | |
404
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
744 #if defined(IA32) |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
745 if (!phase->C->method()->is_strict()) |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
746 // Can't trust native compilers to properly fold strict double |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
747 // division with round-to-zero on this platform. |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
748 #endif |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
749 { |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
750 // If divisor is a constant and not zero, divide them numbers |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
751 if( t1->base() == Type::DoubleCon && |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
752 t2->base() == Type::DoubleCon && |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
753 t2->getd() != 0.0 ) // could be negative zero |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
754 return TypeD::make( t1->getd()/t2->getd() ); |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
755 } |
0 | 756 |
757 // If the dividend is a constant zero | |
758 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
759 // Test TypeF::ZERO is not sufficient as it could be negative zero | |
760 if( t1 == TypeD::ZERO && !g_isnan(t2->getd()) && t2->getd() != 0.0 ) | |
761 return TypeD::ZERO; | |
762 | |
763 // Otherwise we give up all hope | |
764 return Type::DOUBLE; | |
765 } | |
766 | |
767 | |
768 //------------------------------isA_Copy--------------------------------------- | |
769 // Dividing by self is 1. | |
770 // If the divisor is 1, we are an identity on the dividend. | |
771 Node *DivDNode::Identity( PhaseTransform *phase ) { | |
772 return (phase->type( in(2) ) == TypeD::ONE) ? in(1) : this; | |
773 } | |
774 | |
775 //------------------------------Idealize--------------------------------------- | |
776 Node *DivDNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
777 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
305 | 778 // Don't bother trying to transform a dead node |
779 if( in(0) && in(0)->is_top() ) return NULL; | |
0 | 780 |
781 const Type *t2 = phase->type( in(2) ); | |
782 if( t2 == TypeD::ONE ) // Identity? | |
783 return NULL; // Skip it | |
784 | |
785 const TypeD *td = t2->isa_double_constant(); | |
786 if( !td ) return NULL; | |
787 if( td->base() != Type::DoubleCon ) return NULL; | |
788 | |
789 // Check for out of range values | |
790 if( td->is_nan() || !td->is_finite() ) return NULL; | |
791 | |
792 // Get the value | |
793 double d = td->getd(); | |
794 int exp; | |
795 | |
796 // Only for special case of dividing by a power of 2 | |
797 if( frexp(d, &exp) != 0.5 ) return NULL; | |
798 | |
799 // Limit the range of acceptable exponents | |
800 if( exp < -1021 || exp > 1022 ) return NULL; | |
801 | |
802 // Compute the reciprocal | |
803 double reciprocal = 1.0 / d; | |
804 | |
805 assert( frexp(reciprocal, &exp) == 0.5, "reciprocal should be power of 2" ); | |
806 | |
807 // return multiplication by the reciprocal | |
808 return (new (phase->C, 3) MulDNode(in(1), phase->makecon(TypeD::make(reciprocal)))); | |
809 } | |
810 | |
811 //============================================================================= | |
812 //------------------------------Idealize--------------------------------------- | |
813 Node *ModINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
814 // Check for dead control input | |
305 | 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; | |
0 | 818 |
819 // Get the modulus | |
820 const Type *t = phase->type( in(2) ); | |
821 if( t == Type::TOP ) return NULL; | |
822 const TypeInt *ti = t->is_int(); | |
823 | |
824 // Check for useless control input | |
825 // Check for excluding mod-zero case | |
826 if( in(0) && (ti->_hi < 0 || ti->_lo > 0) ) { | |
827 set_req(0, NULL); // Yank control input | |
828 return this; | |
829 } | |
830 | |
831 // See if we are MOD'ing by 2^k or 2^k-1. | |
832 if( !ti->is_con() ) return NULL; | |
833 jint con = ti->get_con(); | |
834 | |
835 Node *hook = new (phase->C, 1) Node(1); | |
836 | |
837 // First, special check for modulo 2^k-1 | |
838 if( con >= 0 && con < max_jint && is_power_of_2(con+1) ) { | |
839 uint k = exact_log2(con+1); // Extract k | |
840 | |
841 // Basic algorithm by David Detlefs. See fastmod_int.java for gory details. | |
842 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*/}; | |
843 int trip_count = 1; | |
844 if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k]; | |
845 | |
846 // If the unroll factor is not too large, and if conditional moves are | |
847 // ok, then use this case | |
848 if( trip_count <= 5 && ConditionalMoveLimit != 0 ) { | |
849 Node *x = in(1); // Value being mod'd | |
850 Node *divisor = in(2); // Also is mask | |
851 | |
852 hook->init_req(0, x); // Add a use to x to prevent him from dying | |
853 // Generate code to reduce X rapidly to nearly 2^k-1. | |
854 for( int i = 0; i < trip_count; i++ ) { | |
145 | 855 Node *xl = phase->transform( new (phase->C, 3) AndINode(x,divisor) ); |
856 Node *xh = phase->transform( new (phase->C, 3) RShiftINode(x,phase->intcon(k)) ); // Must be signed | |
857 x = phase->transform( new (phase->C, 3) AddINode(xh,xl) ); | |
858 hook->set_req(0, x); | |
0 | 859 } |
860 | |
861 // Generate sign-fixup code. Was original value positive? | |
862 // int hack_res = (i >= 0) ? divisor : 1; | |
863 Node *cmp1 = phase->transform( new (phase->C, 3) CmpINode( in(1), phase->intcon(0) ) ); | |
864 Node *bol1 = phase->transform( new (phase->C, 2) BoolNode( cmp1, BoolTest::ge ) ); | |
865 Node *cmov1= phase->transform( new (phase->C, 4) CMoveINode(bol1, phase->intcon(1), divisor, TypeInt::POS) ); | |
866 // if( x >= hack_res ) x -= divisor; | |
867 Node *sub = phase->transform( new (phase->C, 3) SubINode( x, divisor ) ); | |
868 Node *cmp2 = phase->transform( new (phase->C, 3) CmpINode( x, cmov1 ) ); | |
869 Node *bol2 = phase->transform( new (phase->C, 2) BoolNode( cmp2, BoolTest::ge ) ); | |
870 // Convention is to not transform the return value of an Ideal | |
871 // since Ideal is expected to return a modified 'this' or a new node. | |
872 Node *cmov2= new (phase->C, 4) CMoveINode(bol2, x, sub, TypeInt::INT); | |
873 // cmov2 is now the mod | |
874 | |
875 // Now remove the bogus extra edges used to keep things alive | |
876 if (can_reshape) { | |
877 phase->is_IterGVN()->remove_dead_node(hook); | |
878 } else { | |
879 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
880 } | |
881 return cmov2; | |
882 } | |
883 } | |
884 | |
885 // Fell thru, the unroll case is not appropriate. Transform the modulo | |
886 // into a long multiply/int multiply/subtract case | |
887 | |
888 // Cannot handle mod 0, and min_jint isn't handled by the transform | |
889 if( con == 0 || con == min_jint ) return NULL; | |
890 | |
891 // Get the absolute value of the constant; at this point, we can use this | |
892 jint pos_con = (con >= 0) ? con : -con; | |
893 | |
894 // integer Mod 1 is always 0 | |
895 if( pos_con == 1 ) return new (phase->C, 1) ConINode(TypeInt::ZERO); | |
896 | |
897 int log2_con = -1; | |
898 | |
899 // If this is a power of two, they maybe we can mask it | |
900 if( is_power_of_2(pos_con) ) { | |
901 log2_con = log2_intptr((intptr_t)pos_con); | |
902 | |
903 const Type *dt = phase->type(in(1)); | |
904 const TypeInt *dti = dt->isa_int(); | |
905 | |
906 // See if this can be masked, if the dividend is non-negative | |
907 if( dti && dti->_lo >= 0 ) | |
908 return ( new (phase->C, 3) AndINode( in(1), phase->intcon( pos_con-1 ) ) ); | |
909 } | |
910 | |
911 // Save in(1) so that it cannot be changed or deleted | |
912 hook->init_req(0, in(1)); | |
913 | |
914 // Divide using the transform from DivI to MulL | |
145 | 915 Node *result = transform_int_divide( phase, in(1), pos_con ); |
916 if (result != NULL) { | |
917 Node *divide = phase->transform(result); | |
0 | 918 |
145 | 919 // Re-multiply, using a shift if this is a power of two |
920 Node *mult = NULL; | |
0 | 921 |
145 | 922 if( log2_con >= 0 ) |
923 mult = phase->transform( new (phase->C, 3) LShiftINode( divide, phase->intcon( log2_con ) ) ); | |
924 else | |
925 mult = phase->transform( new (phase->C, 3) MulINode( divide, phase->intcon( pos_con ) ) ); | |
0 | 926 |
145 | 927 // Finally, subtract the multiplied divided value from the original |
928 result = new (phase->C, 3) SubINode( in(1), mult ); | |
929 } | |
0 | 930 |
931 // Now remove the bogus extra edges used to keep things alive | |
932 if (can_reshape) { | |
933 phase->is_IterGVN()->remove_dead_node(hook); | |
934 } else { | |
935 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
936 } | |
937 | |
938 // return the value | |
939 return result; | |
940 } | |
941 | |
942 //------------------------------Value------------------------------------------ | |
943 const Type *ModINode::Value( PhaseTransform *phase ) const { | |
944 // Either input is TOP ==> the result is TOP | |
945 const Type *t1 = phase->type( in(1) ); | |
946 const Type *t2 = phase->type( in(2) ); | |
947 if( t1 == Type::TOP ) return Type::TOP; | |
948 if( t2 == Type::TOP ) return Type::TOP; | |
949 | |
950 // We always generate the dynamic check for 0. | |
951 // 0 MOD X is 0 | |
952 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; | |
953 // X MOD X is 0 | |
954 if( phase->eqv( in(1), in(2) ) ) return TypeInt::ZERO; | |
955 | |
956 // Either input is BOTTOM ==> the result is the local BOTTOM | |
957 const Type *bot = bottom_type(); | |
958 if( (t1 == bot) || (t2 == bot) || | |
959 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
960 return bot; | |
961 | |
962 const TypeInt *i1 = t1->is_int(); | |
963 const TypeInt *i2 = t2->is_int(); | |
964 if( !i1->is_con() || !i2->is_con() ) { | |
965 if( i1->_lo >= 0 && i2->_lo >= 0 ) | |
966 return TypeInt::POS; | |
967 // If both numbers are not constants, we know little. | |
968 return TypeInt::INT; | |
969 } | |
970 // Mod by zero? Throw exception at runtime! | |
971 if( !i2->get_con() ) return TypeInt::POS; | |
972 | |
973 // We must be modulo'ing 2 float constants. | |
974 // Check for min_jint % '-1', result is defined to be '0'. | |
975 if( i1->get_con() == min_jint && i2->get_con() == -1 ) | |
976 return TypeInt::ZERO; | |
977 | |
978 return TypeInt::make( i1->get_con() % i2->get_con() ); | |
979 } | |
980 | |
981 | |
982 //============================================================================= | |
983 //------------------------------Idealize--------------------------------------- | |
984 Node *ModLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
985 // Check for dead control input | |
305 | 986 if( in(0) && remove_dead_region(phase, can_reshape) ) return this; |
987 // Don't bother trying to transform a dead node | |
988 if( in(0) && in(0)->is_top() ) return NULL; | |
0 | 989 |
990 // Get the modulus | |
991 const Type *t = phase->type( in(2) ); | |
992 if( t == Type::TOP ) return NULL; | |
145 | 993 const TypeLong *tl = t->is_long(); |
0 | 994 |
995 // Check for useless control input | |
996 // Check for excluding mod-zero case | |
145 | 997 if( in(0) && (tl->_hi < 0 || tl->_lo > 0) ) { |
0 | 998 set_req(0, NULL); // Yank control input |
999 return this; | |
1000 } | |
1001 | |
1002 // See if we are MOD'ing by 2^k or 2^k-1. | |
145 | 1003 if( !tl->is_con() ) return NULL; |
1004 jlong con = tl->get_con(); | |
1005 | |
1006 Node *hook = new (phase->C, 1) Node(1); | |
0 | 1007 |
1008 // Expand mod | |
145 | 1009 if( con >= 0 && con < max_jlong && is_power_of_2_long(con+1) ) { |
568
30663ca5e8f4
6805724: ModLNode::Ideal() generates functionally incorrect graph when divisor is any (2^k-1) constant.
twisti
parents:
567
diff
changeset
|
1010 uint k = exact_log2_long(con+1); // Extract k |
145 | 1011 |
0 | 1012 // Basic algorithm by David Detlefs. See fastmod_long.java for gory details. |
1013 // Used to help a popular random number generator which does a long-mod | |
1014 // of 2^31-1 and shows up in SpecJBB and SciMark. | |
1015 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*/}; | |
1016 int trip_count = 1; | |
1017 if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k]; | |
1018 | |
145 | 1019 // If the unroll factor is not too large, and if conditional moves are |
1020 // ok, then use this case | |
1021 if( trip_count <= 5 && ConditionalMoveLimit != 0 ) { | |
1022 Node *x = in(1); // Value being mod'd | |
1023 Node *divisor = in(2); // Also is mask | |
0 | 1024 |
145 | 1025 hook->init_req(0, x); // Add a use to x to prevent him from dying |
1026 // Generate code to reduce X rapidly to nearly 2^k-1. | |
1027 for( int i = 0; i < trip_count; i++ ) { | |
0 | 1028 Node *xl = phase->transform( new (phase->C, 3) AndLNode(x,divisor) ); |
1029 Node *xh = phase->transform( new (phase->C, 3) RShiftLNode(x,phase->intcon(k)) ); // Must be signed | |
1030 x = phase->transform( new (phase->C, 3) AddLNode(xh,xl) ); | |
1031 hook->set_req(0, x); // Add a use to x to prevent him from dying | |
145 | 1032 } |
1033 | |
1034 // Generate sign-fixup code. Was original value positive? | |
1035 // long hack_res = (i >= 0) ? divisor : CONST64(1); | |
1036 Node *cmp1 = phase->transform( new (phase->C, 3) CmpLNode( in(1), phase->longcon(0) ) ); | |
1037 Node *bol1 = phase->transform( new (phase->C, 2) BoolNode( cmp1, BoolTest::ge ) ); | |
1038 Node *cmov1= phase->transform( new (phase->C, 4) CMoveLNode(bol1, phase->longcon(1), divisor, TypeLong::LONG) ); | |
1039 // if( x >= hack_res ) x -= divisor; | |
1040 Node *sub = phase->transform( new (phase->C, 3) SubLNode( x, divisor ) ); | |
1041 Node *cmp2 = phase->transform( new (phase->C, 3) CmpLNode( x, cmov1 ) ); | |
1042 Node *bol2 = phase->transform( new (phase->C, 2) BoolNode( cmp2, BoolTest::ge ) ); | |
1043 // Convention is to not transform the return value of an Ideal | |
1044 // since Ideal is expected to return a modified 'this' or a new node. | |
1045 Node *cmov2= new (phase->C, 4) CMoveLNode(bol2, x, sub, TypeLong::LONG); | |
1046 // cmov2 is now the mod | |
1047 | |
1048 // Now remove the bogus extra edges used to keep things alive | |
1049 if (can_reshape) { | |
1050 phase->is_IterGVN()->remove_dead_node(hook); | |
1051 } else { | |
1052 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
1053 } | |
1054 return cmov2; | |
0 | 1055 } |
145 | 1056 } |
1057 | |
1058 // Fell thru, the unroll case is not appropriate. Transform the modulo | |
1059 // into a long multiply/int multiply/subtract case | |
1060 | |
1061 // Cannot handle mod 0, and min_jint isn't handled by the transform | |
1062 if( con == 0 || con == min_jlong ) return NULL; | |
1063 | |
1064 // Get the absolute value of the constant; at this point, we can use this | |
1065 jlong pos_con = (con >= 0) ? con : -con; | |
1066 | |
1067 // integer Mod 1 is always 0 | |
1068 if( pos_con == 1 ) return new (phase->C, 1) ConLNode(TypeLong::ZERO); | |
1069 | |
1070 int log2_con = -1; | |
1071 | |
605 | 1072 // If this is a power of two, then maybe we can mask it |
145 | 1073 if( is_power_of_2_long(pos_con) ) { |
1074 log2_con = log2_long(pos_con); | |
1075 | |
1076 const Type *dt = phase->type(in(1)); | |
1077 const TypeLong *dtl = dt->isa_long(); | |
1078 | |
1079 // See if this can be masked, if the dividend is non-negative | |
1080 if( dtl && dtl->_lo >= 0 ) | |
1081 return ( new (phase->C, 3) AndLNode( in(1), phase->longcon( pos_con-1 ) ) ); | |
1082 } | |
0 | 1083 |
145 | 1084 // Save in(1) so that it cannot be changed or deleted |
1085 hook->init_req(0, in(1)); | |
1086 | |
1087 // Divide using the transform from DivI to MulL | |
1088 Node *result = transform_long_divide( phase, in(1), pos_con ); | |
1089 if (result != NULL) { | |
1090 Node *divide = phase->transform(result); | |
1091 | |
1092 // Re-multiply, using a shift if this is a power of two | |
1093 Node *mult = NULL; | |
1094 | |
1095 if( log2_con >= 0 ) | |
1096 mult = phase->transform( new (phase->C, 3) LShiftLNode( divide, phase->intcon( log2_con ) ) ); | |
1097 else | |
1098 mult = phase->transform( new (phase->C, 3) MulLNode( divide, phase->longcon( pos_con ) ) ); | |
1099 | |
1100 // Finally, subtract the multiplied divided value from the original | |
1101 result = new (phase->C, 3) SubLNode( in(1), mult ); | |
0 | 1102 } |
145 | 1103 |
1104 // Now remove the bogus extra edges used to keep things alive | |
1105 if (can_reshape) { | |
1106 phase->is_IterGVN()->remove_dead_node(hook); | |
1107 } else { | |
1108 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
1109 } | |
1110 | |
1111 // return the value | |
1112 return result; | |
0 | 1113 } |
1114 | |
1115 //------------------------------Value------------------------------------------ | |
1116 const Type *ModLNode::Value( PhaseTransform *phase ) const { | |
1117 // Either input is TOP ==> the result is TOP | |
1118 const Type *t1 = phase->type( in(1) ); | |
1119 const Type *t2 = phase->type( in(2) ); | |
1120 if( t1 == Type::TOP ) return Type::TOP; | |
1121 if( t2 == Type::TOP ) return Type::TOP; | |
1122 | |
1123 // We always generate the dynamic check for 0. | |
1124 // 0 MOD X is 0 | |
1125 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; | |
1126 // X MOD X is 0 | |
1127 if( phase->eqv( in(1), in(2) ) ) return TypeLong::ZERO; | |
1128 | |
1129 // Either input is BOTTOM ==> the result is the local BOTTOM | |
1130 const Type *bot = bottom_type(); | |
1131 if( (t1 == bot) || (t2 == bot) || | |
1132 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
1133 return bot; | |
1134 | |
1135 const TypeLong *i1 = t1->is_long(); | |
1136 const TypeLong *i2 = t2->is_long(); | |
1137 if( !i1->is_con() || !i2->is_con() ) { | |
1138 if( i1->_lo >= CONST64(0) && i2->_lo >= CONST64(0) ) | |
1139 return TypeLong::POS; | |
1140 // If both numbers are not constants, we know little. | |
1141 return TypeLong::LONG; | |
1142 } | |
1143 // Mod by zero? Throw exception at runtime! | |
1144 if( !i2->get_con() ) return TypeLong::POS; | |
1145 | |
1146 // We must be modulo'ing 2 float constants. | |
1147 // Check for min_jint % '-1', result is defined to be '0'. | |
1148 if( i1->get_con() == min_jlong && i2->get_con() == -1 ) | |
1149 return TypeLong::ZERO; | |
1150 | |
1151 return TypeLong::make( i1->get_con() % i2->get_con() ); | |
1152 } | |
1153 | |
1154 | |
1155 //============================================================================= | |
1156 //------------------------------Value------------------------------------------ | |
1157 const Type *ModFNode::Value( PhaseTransform *phase ) const { | |
1158 // Either input is TOP ==> the result is TOP | |
1159 const Type *t1 = phase->type( in(1) ); | |
1160 const Type *t2 = phase->type( in(2) ); | |
1161 if( t1 == Type::TOP ) return Type::TOP; | |
1162 if( t2 == Type::TOP ) return Type::TOP; | |
1163 | |
1164 // Either input is BOTTOM ==> the result is the local BOTTOM | |
1165 const Type *bot = bottom_type(); | |
1166 if( (t1 == bot) || (t2 == bot) || | |
1167 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
1168 return bot; | |
1169 | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1170 // 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
|
1171 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
|
1172 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
|
1173 } |
0 | 1174 |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1175 float f1 = t1->getf(); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1176 float f2 = t2->getf(); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1177 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
|
1178 jint x2 = jint_cast(f2); |
0 | 1179 |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1180 // 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
|
1181 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
|
1182 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
|
1183 |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1184 // 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
|
1185 if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jint) |
0 | 1186 return Type::FLOAT; |
1187 | |
1188 // We must be modulo'ing 2 float constants. | |
1189 // 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
|
1190 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
|
1191 if ((x1 ^ xr) < 0) { |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1192 xr ^= min_jint; |
0 | 1193 } |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1194 |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1195 return TypeF::make(jfloat_cast(xr)); |
0 | 1196 } |
1197 | |
1198 | |
1199 //============================================================================= | |
1200 //------------------------------Value------------------------------------------ | |
1201 const Type *ModDNode::Value( PhaseTransform *phase ) const { | |
1202 // Either input is TOP ==> the result is TOP | |
1203 const Type *t1 = phase->type( in(1) ); | |
1204 const Type *t2 = phase->type( in(2) ); | |
1205 if( t1 == Type::TOP ) return Type::TOP; | |
1206 if( t2 == Type::TOP ) return Type::TOP; | |
1207 | |
1208 // Either input is BOTTOM ==> the result is the local BOTTOM | |
1209 const Type *bot = bottom_type(); | |
1210 if( (t1 == bot) || (t2 == bot) || | |
1211 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
1212 return bot; | |
1213 | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1214 // 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
|
1215 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
|
1216 return Type::DOUBLE; // note: x%x can be either NaN or 0 |
0 | 1217 } |
1218 | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1219 double f1 = t1->getd(); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1220 double f2 = t2->getd(); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1221 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
|
1222 jlong x2 = jlong_cast(f2); |
0 | 1223 |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1224 // 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
|
1225 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
|
1226 if (g_isnan(f2)) return t2; |
0 | 1227 |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1228 // 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
|
1229 if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jlong) |
0 | 1230 return Type::DOUBLE; |
1231 | |
1232 // 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
|
1233 // 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
|
1234 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
|
1235 if ((x1 ^ xr) < 0) { |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1236 xr ^= min_jlong; |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1237 } |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1238 |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1239 return TypeD::make(jdouble_cast(xr)); |
0 | 1240 } |
1241 | |
1242 //============================================================================= | |
1243 | |
1244 DivModNode::DivModNode( Node *c, Node *dividend, Node *divisor ) : MultiNode(3) { | |
1245 init_req(0, c); | |
1246 init_req(1, dividend); | |
1247 init_req(2, divisor); | |
1248 } | |
1249 | |
1250 //------------------------------make------------------------------------------ | |
1251 DivModINode* DivModINode::make(Compile* C, Node* div_or_mod) { | |
1252 Node* n = div_or_mod; | |
1253 assert(n->Opcode() == Op_DivI || n->Opcode() == Op_ModI, | |
1254 "only div or mod input pattern accepted"); | |
1255 | |
1256 DivModINode* divmod = new (C, 3) DivModINode(n->in(0), n->in(1), n->in(2)); | |
1257 Node* dproj = new (C, 1) ProjNode(divmod, DivModNode::div_proj_num); | |
1258 Node* mproj = new (C, 1) ProjNode(divmod, DivModNode::mod_proj_num); | |
1259 return divmod; | |
1260 } | |
1261 | |
1262 //------------------------------make------------------------------------------ | |
1263 DivModLNode* DivModLNode::make(Compile* C, Node* div_or_mod) { | |
1264 Node* n = div_or_mod; | |
1265 assert(n->Opcode() == Op_DivL || n->Opcode() == Op_ModL, | |
1266 "only div or mod input pattern accepted"); | |
1267 | |
1268 DivModLNode* divmod = new (C, 3) DivModLNode(n->in(0), n->in(1), n->in(2)); | |
1269 Node* dproj = new (C, 1) ProjNode(divmod, DivModNode::div_proj_num); | |
1270 Node* mproj = new (C, 1) ProjNode(divmod, DivModNode::mod_proj_num); | |
1271 return divmod; | |
1272 } | |
1273 | |
1274 //------------------------------match------------------------------------------ | |
1275 // return result(s) along with their RegMask info | |
1276 Node *DivModINode::match( const ProjNode *proj, const Matcher *match ) { | |
1277 uint ideal_reg = proj->ideal_reg(); | |
1278 RegMask rm; | |
1279 if (proj->_con == div_proj_num) { | |
1280 rm = match->divI_proj_mask(); | |
1281 } else { | |
1282 assert(proj->_con == mod_proj_num, "must be div or mod projection"); | |
1283 rm = match->modI_proj_mask(); | |
1284 } | |
1285 return new (match->C, 1)MachProjNode(this, proj->_con, rm, ideal_reg); | |
1286 } | |
1287 | |
1288 | |
1289 //------------------------------match------------------------------------------ | |
1290 // return result(s) along with their RegMask info | |
1291 Node *DivModLNode::match( const ProjNode *proj, const Matcher *match ) { | |
1292 uint ideal_reg = proj->ideal_reg(); | |
1293 RegMask rm; | |
1294 if (proj->_con == div_proj_num) { | |
1295 rm = match->divL_proj_mask(); | |
1296 } else { | |
1297 assert(proj->_con == mod_proj_num, "must be div or mod projection"); | |
1298 rm = match->modL_proj_mask(); | |
1299 } | |
1300 return new (match->C, 1)MachProjNode(this, proj->_con, rm, ideal_reg); | |
1301 } |