Mercurial > hg > truffle
annotate src/share/vm/opto/divnode.cpp @ 1552:c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
Summary: Change all the Sun copyrights to Oracle copyright
Reviewed-by: ohair
author | trims |
---|---|
date | Thu, 27 May 2010 19:08:38 -0700 |
parents | 174ade00803b |
children | ae065c367d93 |
rev | line source |
---|---|
0 | 1 /* |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1154
diff
changeset
|
2 * Copyright (c) 1997, 2009, Oracle and/or its affiliates. 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 * | |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1154
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1154
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1154
diff
changeset
|
21 * questions. |
0 | 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 ) { |
1154
174ade00803b
6910484: incorrect integer optimization (loosing and op-r in a given example)
kvn
parents:
756
diff
changeset
|
117 if( (-andconi) == d ) // Remove AND if it clears bits which will be shifted |
174ade00803b
6910484: incorrect integer optimization (loosing and op-r in a given example)
kvn
parents:
756
diff
changeset
|
118 dividend = dividend->in(1); |
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
119 needs_rounding = false; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
120 } |
0 | 121 } |
122 } | |
123 | |
124 // Add rounding to the shift to handle the sign bit | |
145 | 125 int l = log2_intptr(d-1)+1; |
126 if (needs_rounding) { | |
127 // Divide-by-power-of-2 can be made into a shift, but you have to do | |
128 // more math for the rounding. You need to add 0 for positive | |
129 // numbers, and "i-1" for negative numbers. Example: i=4, so the | |
130 // shift is by 2. You need to add 3 to negative dividends and 0 to | |
131 // positive ones. So (-7+3)>>2 becomes -1, (-4+3)>>2 becomes -1, | |
132 // (-2+3)>>2 becomes 0, etc. | |
133 | |
134 // Compute 0 or -1, based on sign bit | |
135 Node *sign = phase->transform(new (phase->C, 3) RShiftINode(dividend, phase->intcon(N - 1))); | |
136 // Mask sign bit to the low sign bits | |
137 Node *round = phase->transform(new (phase->C, 3) URShiftINode(sign, phase->intcon(N - l))); | |
138 // Round up before shifting | |
139 dividend = phase->transform(new (phase->C, 3) AddINode(dividend, round)); | |
0 | 140 } |
141 | |
145 | 142 // Shift for division |
0 | 143 q = new (phase->C, 3) RShiftINode(dividend, phase->intcon(l)); |
144 | |
145 | 145 if (!d_pos) { |
0 | 146 q = new (phase->C, 3) SubINode(phase->intcon(0), phase->transform(q)); |
145 | 147 } |
148 } else { | |
149 // Attempt the jint constant divide -> multiply transform found in | |
150 // "Division by Invariant Integers using Multiplication" | |
151 // by Granlund and Montgomery | |
152 // See also "Hacker's Delight", chapter 10 by Warren. | |
153 | |
154 jint magic_const; | |
155 jint shift_const; | |
156 if (magic_int_divide_constants(d, magic_const, shift_const)) { | |
157 Node *magic = phase->longcon(magic_const); | |
158 Node *dividend_long = phase->transform(new (phase->C, 2) ConvI2LNode(dividend)); | |
159 | |
160 // Compute the high half of the dividend x magic multiplication | |
161 Node *mul_hi = phase->transform(new (phase->C, 3) MulLNode(dividend_long, magic)); | |
162 | |
163 if (magic_const < 0) { | |
164 mul_hi = phase->transform(new (phase->C, 3) RShiftLNode(mul_hi, phase->intcon(N))); | |
165 mul_hi = phase->transform(new (phase->C, 2) ConvL2INode(mul_hi)); | |
166 | |
167 // The magic multiplier is too large for a 32 bit constant. We've adjusted | |
168 // it down by 2^32, but have to add 1 dividend back in after the multiplication. | |
169 // This handles the "overflow" case described by Granlund and Montgomery. | |
170 mul_hi = phase->transform(new (phase->C, 3) AddINode(dividend, mul_hi)); | |
171 | |
172 // Shift over the (adjusted) mulhi | |
173 if (shift_const != 0) { | |
174 mul_hi = phase->transform(new (phase->C, 3) RShiftINode(mul_hi, phase->intcon(shift_const))); | |
175 } | |
176 } else { | |
177 // No add is required, we can merge the shifts together. | |
178 mul_hi = phase->transform(new (phase->C, 3) RShiftLNode(mul_hi, phase->intcon(N + shift_const))); | |
179 mul_hi = phase->transform(new (phase->C, 2) ConvL2INode(mul_hi)); | |
180 } | |
181 | |
182 // Get a 0 or -1 from the sign of the dividend. | |
183 Node *addend0 = mul_hi; | |
184 Node *addend1 = phase->transform(new (phase->C, 3) RShiftINode(dividend, phase->intcon(N-1))); | |
185 | |
186 // If the divisor is negative, swap the order of the input addends; | |
187 // this has the effect of negating the quotient. | |
188 if (!d_pos) { | |
189 Node *temp = addend0; addend0 = addend1; addend1 = temp; | |
190 } | |
191 | |
192 // Adjust the final quotient by subtracting -1 (adding 1) | |
193 // from the mul_hi. | |
194 q = new (phase->C, 3) SubINode(addend0, addend1); | |
195 } | |
196 } | |
197 | |
198 return q; | |
199 } | |
200 | |
201 //---------------------magic_long_divide_constants----------------------------- | |
202 // Compute magic multiplier and shift constant for converting a 64 bit divide | |
203 // by constant into a multiply/shift/add series. Return false if calculations | |
204 // fail. | |
205 // | |
605 | 206 // Borrowed almost verbatim from Hacker's Delight by Henry S. Warren, Jr. with |
145 | 207 // minor type name and parameter changes. Adjusted to 64 bit word width. |
208 static bool magic_long_divide_constants(jlong d, jlong &M, jint &s) { | |
209 int64_t p; | |
210 uint64_t ad, anc, delta, q1, r1, q2, r2, t; | |
211 const uint64_t two63 = 0x8000000000000000LL; // 2**63. | |
212 | |
213 ad = ABS(d); | |
214 if (d == 0 || d == 1) return false; | |
215 t = two63 + ((uint64_t)d >> 63); | |
216 anc = t - 1 - t%ad; // Absolute value of nc. | |
217 p = 63; // Init. p. | |
218 q1 = two63/anc; // Init. q1 = 2**p/|nc|. | |
219 r1 = two63 - q1*anc; // Init. r1 = rem(2**p, |nc|). | |
220 q2 = two63/ad; // Init. q2 = 2**p/|d|. | |
221 r2 = two63 - q2*ad; // Init. r2 = rem(2**p, |d|). | |
222 do { | |
223 p = p + 1; | |
224 q1 = 2*q1; // Update q1 = 2**p/|nc|. | |
225 r1 = 2*r1; // Update r1 = rem(2**p, |nc|). | |
226 if (r1 >= anc) { // (Must be an unsigned | |
227 q1 = q1 + 1; // comparison here). | |
228 r1 = r1 - anc; | |
229 } | |
230 q2 = 2*q2; // Update q2 = 2**p/|d|. | |
231 r2 = 2*r2; // Update r2 = rem(2**p, |d|). | |
232 if (r2 >= ad) { // (Must be an unsigned | |
233 q2 = q2 + 1; // comparison here). | |
234 r2 = r2 - ad; | |
235 } | |
236 delta = ad - r2; | |
237 } while (q1 < delta || (q1 == delta && r1 == 0)); | |
238 | |
239 M = q2 + 1; | |
240 if (d < 0) M = -M; // Magic number and | |
241 s = p - 64; // shift amount to return. | |
242 | |
243 return true; | |
244 } | |
245 | |
246 //---------------------long_by_long_mulhi-------------------------------------- | |
247 // 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
|
248 static Node* long_by_long_mulhi(PhaseGVN* phase, Node* dividend, jlong magic_const) { |
145 | 249 // If the architecture supports a 64x64 mulhi, there is |
250 // no need to synthesize it in ideal nodes. | |
251 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
|
252 Node* v = phase->longcon(magic_const); |
145 | 253 return new (phase->C, 3) MulHiLNode(dividend, v); |
0 | 254 } |
255 | |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
256 // 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
|
257 // (http://www.hackersdelight.org/HDcode/mulhs.c) |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
258 // |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
259 // int mulhs(int u, int v) { |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
260 // unsigned u0, v0, w0; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
261 // int u1, v1, w1, w2, t; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
262 // |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
263 // u0 = u & 0xFFFF; u1 = u >> 16; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
264 // v0 = v & 0xFFFF; v1 = v >> 16; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
265 // w0 = u0*v0; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
266 // t = u1*v0 + (w0 >> 16); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
267 // w1 = t & 0xFFFF; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
268 // w2 = t >> 16; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
269 // w1 = u0*v1 + w1; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
270 // return u1*v1 + w2 + (w1 >> 16); |
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 // |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
273 // 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
|
274 // following inline comments are adapted to 64x64. |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
275 |
145 | 276 const int N = 64; |
277 | |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
278 // u0 = u & 0xFFFFFFFF; u1 = u >> 32; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
279 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
|
280 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
|
281 |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
282 // v0 = v & 0xFFFFFFFF; v1 = v >> 32; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
283 Node* v0 = phase->longcon(magic_const & 0xFFFFFFFF); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
284 Node* v1 = phase->longcon(magic_const >> (N / 2)); |
145 | 285 |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
286 // w0 = u0*v0; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
287 Node* w0 = phase->transform(new (phase->C, 3) MulLNode(u0, v0)); |
145 | 288 |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
289 // t = u1*v0 + (w0 >> 32); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
290 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
|
291 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
|
292 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
|
293 |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
294 // w1 = t & 0xFFFFFFFF; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
295 Node* w1 = new (phase->C, 3) AndLNode(t, phase->longcon(0xFFFFFFFF)); |
145 | 296 |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
297 // w2 = t >> 32; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
298 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
|
299 |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
300 // 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
|
301 // doesn't go dead prematurely. |
756 | 302 // 6837011: We need to transform w2 before w1 because the |
303 // transformation of w1 could return t. | |
304 w2 = phase->transform(w2); | |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
305 w1 = phase->transform(w1); |
294
616a07a75c3c
6732154: REG: Printing an Image using image/gif doc flavor crashes the VM, Solsparc
rasbold
parents:
196
diff
changeset
|
306 |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
307 // w1 = u0*v1 + w1; |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
308 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
|
309 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
|
310 |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
311 // return u1*v1 + w2 + (w1 >> 32); |
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
312 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
|
313 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
|
314 Node* temp2 = phase->transform(new (phase->C, 3) RShiftLNode(w1, phase->intcon(N / 2))); |
145 | 315 |
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
316 return new (phase->C, 3) AddLNode(temp1, temp2); |
145 | 317 } |
318 | |
319 | |
320 //--------------------------transform_long_divide------------------------------ | |
321 // Convert a division by constant divisor into an alternate Ideal graph. | |
322 // Return NULL if no transformation occurs. | |
323 static Node *transform_long_divide( PhaseGVN *phase, Node *dividend, jlong divisor ) { | |
324 // Check for invalid divisors | |
325 assert( divisor != 0L && divisor != min_jlong, | |
326 "bad divisor for transforming to long multiply" ); | |
327 | |
328 bool d_pos = divisor >= 0; | |
329 jlong d = d_pos ? divisor : -divisor; | |
330 const int N = 64; | |
331 | |
332 // Result | |
333 Node *q = NULL; | |
334 | |
335 if (d == 1) { | |
336 // division by +/- 1 | |
337 if (!d_pos) { | |
338 // Just negate the value | |
339 q = new (phase->C, 3) SubLNode(phase->longcon(0), dividend); | |
340 } | |
341 } else if ( is_power_of_2_long(d) ) { | |
342 | |
343 // division by +/- a power of 2 | |
344 | |
345 // See if we can simply do a shift without rounding | |
346 bool needs_rounding = true; | |
347 const Type *dt = phase->type(dividend); | |
348 const TypeLong *dtl = dt->isa_long(); | |
0 | 349 |
145 | 350 if (dtl && dtl->_lo > 0) { |
351 // we don't need to round a positive dividend | |
352 needs_rounding = false; | |
353 } else if( dividend->Opcode() == Op_AndL ) { | |
354 // An AND mask of sufficient size clears the low bits and | |
355 // I can avoid rounding. | |
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
356 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
|
357 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
|
358 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
|
359 if( andconl < 0 && is_power_of_2_long(-andconl) && (-andconl) >= d ) { |
1154
174ade00803b
6910484: incorrect integer optimization (loosing and op-r in a given example)
kvn
parents:
756
diff
changeset
|
360 if( (-andconl) == d ) // Remove AND if it clears bits which will be shifted |
174ade00803b
6910484: incorrect integer optimization (loosing and op-r in a given example)
kvn
parents:
756
diff
changeset
|
361 dividend = dividend->in(1); |
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
362 needs_rounding = false; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
363 } |
145 | 364 } |
365 } | |
366 | |
367 // Add rounding to the shift to handle the sign bit | |
368 int l = log2_long(d-1)+1; | |
369 if (needs_rounding) { | |
370 // Divide-by-power-of-2 can be made into a shift, but you have to do | |
371 // more math for the rounding. You need to add 0 for positive | |
372 // numbers, and "i-1" for negative numbers. Example: i=4, so the | |
373 // shift is by 2. You need to add 3 to negative dividends and 0 to | |
374 // positive ones. So (-7+3)>>2 becomes -1, (-4+3)>>2 becomes -1, | |
375 // (-2+3)>>2 becomes 0, etc. | |
376 | |
377 // Compute 0 or -1, based on sign bit | |
378 Node *sign = phase->transform(new (phase->C, 3) RShiftLNode(dividend, phase->intcon(N - 1))); | |
379 // Mask sign bit to the low sign bits | |
380 Node *round = phase->transform(new (phase->C, 3) URShiftLNode(sign, phase->intcon(N - l))); | |
381 // Round up before shifting | |
382 dividend = phase->transform(new (phase->C, 3) AddLNode(dividend, round)); | |
383 } | |
384 | |
385 // Shift for division | |
386 q = new (phase->C, 3) RShiftLNode(dividend, phase->intcon(l)); | |
387 | |
388 if (!d_pos) { | |
389 q = new (phase->C, 3) SubLNode(phase->longcon(0), phase->transform(q)); | |
390 } | |
391 } else { | |
392 // Attempt the jlong constant divide -> multiply transform found in | |
393 // "Division by Invariant Integers using Multiplication" | |
394 // by Granlund and Montgomery | |
395 // See also "Hacker's Delight", chapter 10 by Warren. | |
396 | |
397 jlong magic_const; | |
398 jint shift_const; | |
399 if (magic_long_divide_constants(d, magic_const, shift_const)) { | |
400 // Compute the high half of the dividend x magic multiplication | |
401 Node *mul_hi = phase->transform(long_by_long_mulhi(phase, dividend, magic_const)); | |
402 | |
403 // The high half of the 128-bit multiply is computed. | |
404 if (magic_const < 0) { | |
405 // The magic multiplier is too large for a 64 bit constant. We've adjusted | |
406 // it down by 2^64, but have to add 1 dividend back in after the multiplication. | |
407 // This handles the "overflow" case described by Granlund and Montgomery. | |
408 mul_hi = phase->transform(new (phase->C, 3) AddLNode(dividend, mul_hi)); | |
409 } | |
410 | |
411 // Shift over the (adjusted) mulhi | |
412 if (shift_const != 0) { | |
413 mul_hi = phase->transform(new (phase->C, 3) RShiftLNode(mul_hi, phase->intcon(shift_const))); | |
414 } | |
415 | |
416 // Get a 0 or -1 from the sign of the dividend. | |
417 Node *addend0 = mul_hi; | |
418 Node *addend1 = phase->transform(new (phase->C, 3) RShiftLNode(dividend, phase->intcon(N-1))); | |
419 | |
420 // If the divisor is negative, swap the order of the input addends; | |
421 // this has the effect of negating the quotient. | |
422 if (!d_pos) { | |
423 Node *temp = addend0; addend0 = addend1; addend1 = temp; | |
424 } | |
425 | |
426 // Adjust the final quotient by subtracting -1 (adding 1) | |
427 // from the mul_hi. | |
428 q = new (phase->C, 3) SubLNode(addend0, addend1); | |
429 } | |
0 | 430 } |
431 | |
145 | 432 return q; |
0 | 433 } |
434 | |
435 //============================================================================= | |
436 //------------------------------Identity--------------------------------------- | |
437 // If the divisor is 1, we are an identity on the dividend. | |
438 Node *DivINode::Identity( PhaseTransform *phase ) { | |
439 return (phase->type( in(2) )->higher_equal(TypeInt::ONE)) ? in(1) : this; | |
440 } | |
441 | |
442 //------------------------------Idealize--------------------------------------- | |
443 // Divides can be changed to multiplies and/or shifts | |
444 Node *DivINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
445 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
305 | 446 // Don't bother trying to transform a dead node |
447 if( in(0) && in(0)->is_top() ) return NULL; | |
0 | 448 |
449 const Type *t = phase->type( in(2) ); | |
450 if( t == TypeInt::ONE ) // Identity? | |
451 return NULL; // Skip it | |
452 | |
453 const TypeInt *ti = t->isa_int(); | |
454 if( !ti ) return NULL; | |
455 if( !ti->is_con() ) return NULL; | |
145 | 456 jint i = ti->get_con(); // Get divisor |
0 | 457 |
458 if (i == 0) return NULL; // Dividing by zero constant does not idealize | |
459 | |
460 set_req(0,NULL); // Dividing by a not-zero constant; no faulting | |
461 | |
462 // Dividing by MININT does not optimize as a power-of-2 shift. | |
463 if( i == min_jint ) return NULL; | |
464 | |
145 | 465 return transform_int_divide( phase, in(1), i ); |
0 | 466 } |
467 | |
468 //------------------------------Value------------------------------------------ | |
469 // A DivINode divides its inputs. The third input is a Control input, used to | |
470 // prevent hoisting the divide above an unsafe test. | |
471 const Type *DivINode::Value( PhaseTransform *phase ) const { | |
472 // Either input is TOP ==> the result is TOP | |
473 const Type *t1 = phase->type( in(1) ); | |
474 const Type *t2 = phase->type( in(2) ); | |
475 if( t1 == Type::TOP ) return Type::TOP; | |
476 if( t2 == Type::TOP ) return Type::TOP; | |
477 | |
478 // x/x == 1 since we always generate the dynamic divisor check for 0. | |
479 if( phase->eqv( in(1), in(2) ) ) | |
480 return TypeInt::ONE; | |
481 | |
482 // Either input is BOTTOM ==> the result is the local BOTTOM | |
483 const Type *bot = bottom_type(); | |
484 if( (t1 == bot) || (t2 == bot) || | |
485 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
486 return bot; | |
487 | |
488 // Divide the two numbers. We approximate. | |
489 // If divisor is a constant and not zero | |
490 const TypeInt *i1 = t1->is_int(); | |
491 const TypeInt *i2 = t2->is_int(); | |
492 int widen = MAX2(i1->_widen, i2->_widen); | |
493 | |
494 if( i2->is_con() && i2->get_con() != 0 ) { | |
495 int32 d = i2->get_con(); // Divisor | |
496 jint lo, hi; | |
497 if( d >= 0 ) { | |
498 lo = i1->_lo/d; | |
499 hi = i1->_hi/d; | |
500 } else { | |
501 if( d == -1 && i1->_lo == min_jint ) { | |
502 // 'min_jint/-1' throws arithmetic exception during compilation | |
503 lo = min_jint; | |
504 // do not support holes, 'hi' must go to either min_jint or max_jint: | |
505 // [min_jint, -10]/[-1,-1] ==> [min_jint] UNION [10,max_jint] | |
506 hi = i1->_hi == min_jint ? min_jint : max_jint; | |
507 } else { | |
508 lo = i1->_hi/d; | |
509 hi = i1->_lo/d; | |
510 } | |
511 } | |
512 return TypeInt::make(lo, hi, widen); | |
513 } | |
514 | |
515 // If the dividend is a constant | |
516 if( i1->is_con() ) { | |
517 int32 d = i1->get_con(); | |
518 if( d < 0 ) { | |
519 if( d == min_jint ) { | |
520 // (-min_jint) == min_jint == (min_jint / -1) | |
521 return TypeInt::make(min_jint, max_jint/2 + 1, widen); | |
522 } else { | |
523 return TypeInt::make(d, -d, widen); | |
524 } | |
525 } | |
526 return TypeInt::make(-d, d, widen); | |
527 } | |
528 | |
529 // Otherwise we give up all hope | |
530 return TypeInt::INT; | |
531 } | |
532 | |
533 | |
534 //============================================================================= | |
535 //------------------------------Identity--------------------------------------- | |
536 // If the divisor is 1, we are an identity on the dividend. | |
537 Node *DivLNode::Identity( PhaseTransform *phase ) { | |
538 return (phase->type( in(2) )->higher_equal(TypeLong::ONE)) ? in(1) : this; | |
539 } | |
540 | |
541 //------------------------------Idealize--------------------------------------- | |
542 // Dividing by a power of 2 is a shift. | |
543 Node *DivLNode::Ideal( PhaseGVN *phase, bool can_reshape) { | |
544 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
305 | 545 // Don't bother trying to transform a dead node |
546 if( in(0) && in(0)->is_top() ) return NULL; | |
0 | 547 |
548 const Type *t = phase->type( in(2) ); | |
145 | 549 if( t == TypeLong::ONE ) // Identity? |
0 | 550 return NULL; // Skip it |
551 | |
145 | 552 const TypeLong *tl = t->isa_long(); |
553 if( !tl ) return NULL; | |
554 if( !tl->is_con() ) return NULL; | |
555 jlong l = tl->get_con(); // Get divisor | |
556 | |
557 if (l == 0) return NULL; // Dividing by zero constant does not idealize | |
558 | |
559 set_req(0,NULL); // Dividing by a not-zero constant; no faulting | |
0 | 560 |
561 // Dividing by MININT does not optimize as a power-of-2 shift. | |
145 | 562 if( l == min_jlong ) return NULL; |
0 | 563 |
145 | 564 return transform_long_divide( phase, in(1), l ); |
0 | 565 } |
566 | |
567 //------------------------------Value------------------------------------------ | |
568 // A DivLNode divides its inputs. The third input is a Control input, used to | |
569 // prevent hoisting the divide above an unsafe test. | |
570 const Type *DivLNode::Value( PhaseTransform *phase ) const { | |
571 // Either input is TOP ==> the result is TOP | |
572 const Type *t1 = phase->type( in(1) ); | |
573 const Type *t2 = phase->type( in(2) ); | |
574 if( t1 == Type::TOP ) return Type::TOP; | |
575 if( t2 == Type::TOP ) return Type::TOP; | |
576 | |
577 // x/x == 1 since we always generate the dynamic divisor check for 0. | |
578 if( phase->eqv( in(1), in(2) ) ) | |
579 return TypeLong::ONE; | |
580 | |
581 // Either input is BOTTOM ==> the result is the local BOTTOM | |
582 const Type *bot = bottom_type(); | |
583 if( (t1 == bot) || (t2 == bot) || | |
584 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
585 return bot; | |
586 | |
587 // Divide the two numbers. We approximate. | |
588 // If divisor is a constant and not zero | |
589 const TypeLong *i1 = t1->is_long(); | |
590 const TypeLong *i2 = t2->is_long(); | |
591 int widen = MAX2(i1->_widen, i2->_widen); | |
592 | |
593 if( i2->is_con() && i2->get_con() != 0 ) { | |
594 jlong d = i2->get_con(); // Divisor | |
595 jlong lo, hi; | |
596 if( d >= 0 ) { | |
597 lo = i1->_lo/d; | |
598 hi = i1->_hi/d; | |
599 } else { | |
600 if( d == CONST64(-1) && i1->_lo == min_jlong ) { | |
601 // 'min_jlong/-1' throws arithmetic exception during compilation | |
602 lo = min_jlong; | |
603 // do not support holes, 'hi' must go to either min_jlong or max_jlong: | |
604 // [min_jlong, -10]/[-1,-1] ==> [min_jlong] UNION [10,max_jlong] | |
605 hi = i1->_hi == min_jlong ? min_jlong : max_jlong; | |
606 } else { | |
607 lo = i1->_hi/d; | |
608 hi = i1->_lo/d; | |
609 } | |
610 } | |
611 return TypeLong::make(lo, hi, widen); | |
612 } | |
613 | |
614 // If the dividend is a constant | |
615 if( i1->is_con() ) { | |
616 jlong d = i1->get_con(); | |
617 if( d < 0 ) { | |
618 if( d == min_jlong ) { | |
619 // (-min_jlong) == min_jlong == (min_jlong / -1) | |
620 return TypeLong::make(min_jlong, max_jlong/2 + 1, widen); | |
621 } else { | |
622 return TypeLong::make(d, -d, widen); | |
623 } | |
624 } | |
625 return TypeLong::make(-d, d, widen); | |
626 } | |
627 | |
628 // Otherwise we give up all hope | |
629 return TypeLong::LONG; | |
630 } | |
631 | |
632 | |
633 //============================================================================= | |
634 //------------------------------Value------------------------------------------ | |
635 // An DivFNode divides its inputs. The third input is a Control input, used to | |
636 // prevent hoisting the divide above an unsafe test. | |
637 const Type *DivFNode::Value( PhaseTransform *phase ) const { | |
638 // Either input is TOP ==> the result is TOP | |
639 const Type *t1 = phase->type( in(1) ); | |
640 const Type *t2 = phase->type( in(2) ); | |
641 if( t1 == Type::TOP ) return Type::TOP; | |
642 if( t2 == Type::TOP ) return Type::TOP; | |
643 | |
644 // Either input is BOTTOM ==> the result is the local BOTTOM | |
645 const Type *bot = bottom_type(); | |
646 if( (t1 == bot) || (t2 == bot) || | |
647 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
648 return bot; | |
649 | |
650 // x/x == 1, we ignore 0/0. | |
651 // 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
|
652 // Does not work for variables because of NaN's |
0 | 653 if( phase->eqv( in(1), in(2) ) && t1->base() == Type::FloatCon) |
654 if (!g_isnan(t1->getf()) && g_isfinite(t1->getf()) && t1->getf() != 0.0) // could be negative ZERO or NaN | |
655 return TypeF::ONE; | |
656 | |
657 if( t2 == TypeF::ONE ) | |
658 return t1; | |
659 | |
660 // If divisor is a constant and not zero, divide them numbers | |
661 if( t1->base() == Type::FloatCon && | |
662 t2->base() == Type::FloatCon && | |
663 t2->getf() != 0.0 ) // could be negative zero | |
664 return TypeF::make( t1->getf()/t2->getf() ); | |
665 | |
666 // If the dividend is a constant zero | |
667 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
668 // Test TypeF::ZERO is not sufficient as it could be negative zero | |
669 | |
670 if( t1 == TypeF::ZERO && !g_isnan(t2->getf()) && t2->getf() != 0.0 ) | |
671 return TypeF::ZERO; | |
672 | |
673 // Otherwise we give up all hope | |
674 return Type::FLOAT; | |
675 } | |
676 | |
677 //------------------------------isA_Copy--------------------------------------- | |
678 // Dividing by self is 1. | |
679 // If the divisor is 1, we are an identity on the dividend. | |
680 Node *DivFNode::Identity( PhaseTransform *phase ) { | |
681 return (phase->type( in(2) ) == TypeF::ONE) ? in(1) : this; | |
682 } | |
683 | |
684 | |
685 //------------------------------Idealize--------------------------------------- | |
686 Node *DivFNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
687 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
305 | 688 // Don't bother trying to transform a dead node |
689 if( in(0) && in(0)->is_top() ) return NULL; | |
0 | 690 |
691 const Type *t2 = phase->type( in(2) ); | |
692 if( t2 == TypeF::ONE ) // Identity? | |
693 return NULL; // Skip it | |
694 | |
695 const TypeF *tf = t2->isa_float_constant(); | |
696 if( !tf ) return NULL; | |
697 if( tf->base() != Type::FloatCon ) return NULL; | |
698 | |
699 // Check for out of range values | |
700 if( tf->is_nan() || !tf->is_finite() ) return NULL; | |
701 | |
702 // Get the value | |
703 float f = tf->getf(); | |
704 int exp; | |
705 | |
706 // Only for special case of dividing by a power of 2 | |
707 if( frexp((double)f, &exp) != 0.5 ) return NULL; | |
708 | |
709 // Limit the range of acceptable exponents | |
710 if( exp < -126 || exp > 126 ) return NULL; | |
711 | |
712 // Compute the reciprocal | |
713 float reciprocal = ((float)1.0) / f; | |
714 | |
715 assert( frexp((double)reciprocal, &exp) == 0.5, "reciprocal should be power of 2" ); | |
716 | |
717 // return multiplication by the reciprocal | |
718 return (new (phase->C, 3) MulFNode(in(1), phase->makecon(TypeF::make(reciprocal)))); | |
719 } | |
720 | |
721 //============================================================================= | |
722 //------------------------------Value------------------------------------------ | |
723 // 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
|
724 // prevent hoisting the divide above an unsafe test. |
0 | 725 const Type *DivDNode::Value( PhaseTransform *phase ) const { |
726 // Either input is TOP ==> the result is TOP | |
727 const Type *t1 = phase->type( in(1) ); | |
728 const Type *t2 = phase->type( in(2) ); | |
729 if( t1 == Type::TOP ) return Type::TOP; | |
730 if( t2 == Type::TOP ) return Type::TOP; | |
731 | |
732 // Either input is BOTTOM ==> the result is the local BOTTOM | |
733 const Type *bot = bottom_type(); | |
734 if( (t1 == bot) || (t2 == bot) || | |
735 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
736 return bot; | |
737 | |
738 // x/x == 1, we ignore 0/0. | |
739 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
740 // Does not work for variables because of NaN's | |
741 if( phase->eqv( in(1), in(2) ) && t1->base() == Type::DoubleCon) | |
742 if (!g_isnan(t1->getd()) && g_isfinite(t1->getd()) && t1->getd() != 0.0) // could be negative ZERO or NaN | |
743 return TypeD::ONE; | |
744 | |
745 if( t2 == TypeD::ONE ) | |
746 return t1; | |
747 | |
404
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
748 #if defined(IA32) |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
749 if (!phase->C->method()->is_strict()) |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
750 // 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
|
751 // division with round-to-zero on this platform. |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
752 #endif |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
753 { |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
754 // 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
|
755 if( t1->base() == Type::DoubleCon && |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
756 t2->base() == Type::DoubleCon && |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
757 t2->getd() != 0.0 ) // could be negative zero |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
758 return TypeD::make( t1->getd()/t2->getd() ); |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
759 } |
0 | 760 |
761 // If the dividend is a constant zero | |
762 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
763 // Test TypeF::ZERO is not sufficient as it could be negative zero | |
764 if( t1 == TypeD::ZERO && !g_isnan(t2->getd()) && t2->getd() != 0.0 ) | |
765 return TypeD::ZERO; | |
766 | |
767 // Otherwise we give up all hope | |
768 return Type::DOUBLE; | |
769 } | |
770 | |
771 | |
772 //------------------------------isA_Copy--------------------------------------- | |
773 // Dividing by self is 1. | |
774 // If the divisor is 1, we are an identity on the dividend. | |
775 Node *DivDNode::Identity( PhaseTransform *phase ) { | |
776 return (phase->type( in(2) ) == TypeD::ONE) ? in(1) : this; | |
777 } | |
778 | |
779 //------------------------------Idealize--------------------------------------- | |
780 Node *DivDNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
781 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
305 | 782 // Don't bother trying to transform a dead node |
783 if( in(0) && in(0)->is_top() ) return NULL; | |
0 | 784 |
785 const Type *t2 = phase->type( in(2) ); | |
786 if( t2 == TypeD::ONE ) // Identity? | |
787 return NULL; // Skip it | |
788 | |
789 const TypeD *td = t2->isa_double_constant(); | |
790 if( !td ) return NULL; | |
791 if( td->base() != Type::DoubleCon ) return NULL; | |
792 | |
793 // Check for out of range values | |
794 if( td->is_nan() || !td->is_finite() ) return NULL; | |
795 | |
796 // Get the value | |
797 double d = td->getd(); | |
798 int exp; | |
799 | |
800 // Only for special case of dividing by a power of 2 | |
801 if( frexp(d, &exp) != 0.5 ) return NULL; | |
802 | |
803 // Limit the range of acceptable exponents | |
804 if( exp < -1021 || exp > 1022 ) return NULL; | |
805 | |
806 // Compute the reciprocal | |
807 double reciprocal = 1.0 / d; | |
808 | |
809 assert( frexp(reciprocal, &exp) == 0.5, "reciprocal should be power of 2" ); | |
810 | |
811 // return multiplication by the reciprocal | |
812 return (new (phase->C, 3) MulDNode(in(1), phase->makecon(TypeD::make(reciprocal)))); | |
813 } | |
814 | |
815 //============================================================================= | |
816 //------------------------------Idealize--------------------------------------- | |
817 Node *ModINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
818 // Check for dead control input | |
305 | 819 if( in(0) && remove_dead_region(phase, can_reshape) ) return this; |
820 // Don't bother trying to transform a dead node | |
821 if( in(0) && in(0)->is_top() ) return NULL; | |
0 | 822 |
823 // Get the modulus | |
824 const Type *t = phase->type( in(2) ); | |
825 if( t == Type::TOP ) return NULL; | |
826 const TypeInt *ti = t->is_int(); | |
827 | |
828 // Check for useless control input | |
829 // Check for excluding mod-zero case | |
830 if( in(0) && (ti->_hi < 0 || ti->_lo > 0) ) { | |
831 set_req(0, NULL); // Yank control input | |
832 return this; | |
833 } | |
834 | |
835 // See if we are MOD'ing by 2^k or 2^k-1. | |
836 if( !ti->is_con() ) return NULL; | |
837 jint con = ti->get_con(); | |
838 | |
839 Node *hook = new (phase->C, 1) Node(1); | |
840 | |
841 // First, special check for modulo 2^k-1 | |
842 if( con >= 0 && con < max_jint && is_power_of_2(con+1) ) { | |
843 uint k = exact_log2(con+1); // Extract k | |
844 | |
845 // Basic algorithm by David Detlefs. See fastmod_int.java for gory details. | |
846 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*/}; | |
847 int trip_count = 1; | |
848 if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k]; | |
849 | |
850 // If the unroll factor is not too large, and if conditional moves are | |
851 // ok, then use this case | |
852 if( trip_count <= 5 && ConditionalMoveLimit != 0 ) { | |
853 Node *x = in(1); // Value being mod'd | |
854 Node *divisor = in(2); // Also is mask | |
855 | |
856 hook->init_req(0, x); // Add a use to x to prevent him from dying | |
857 // Generate code to reduce X rapidly to nearly 2^k-1. | |
858 for( int i = 0; i < trip_count; i++ ) { | |
145 | 859 Node *xl = phase->transform( new (phase->C, 3) AndINode(x,divisor) ); |
860 Node *xh = phase->transform( new (phase->C, 3) RShiftINode(x,phase->intcon(k)) ); // Must be signed | |
861 x = phase->transform( new (phase->C, 3) AddINode(xh,xl) ); | |
862 hook->set_req(0, x); | |
0 | 863 } |
864 | |
865 // Generate sign-fixup code. Was original value positive? | |
866 // int hack_res = (i >= 0) ? divisor : 1; | |
867 Node *cmp1 = phase->transform( new (phase->C, 3) CmpINode( in(1), phase->intcon(0) ) ); | |
868 Node *bol1 = phase->transform( new (phase->C, 2) BoolNode( cmp1, BoolTest::ge ) ); | |
869 Node *cmov1= phase->transform( new (phase->C, 4) CMoveINode(bol1, phase->intcon(1), divisor, TypeInt::POS) ); | |
870 // if( x >= hack_res ) x -= divisor; | |
871 Node *sub = phase->transform( new (phase->C, 3) SubINode( x, divisor ) ); | |
872 Node *cmp2 = phase->transform( new (phase->C, 3) CmpINode( x, cmov1 ) ); | |
873 Node *bol2 = phase->transform( new (phase->C, 2) BoolNode( cmp2, BoolTest::ge ) ); | |
874 // Convention is to not transform the return value of an Ideal | |
875 // since Ideal is expected to return a modified 'this' or a new node. | |
876 Node *cmov2= new (phase->C, 4) CMoveINode(bol2, x, sub, TypeInt::INT); | |
877 // cmov2 is now the mod | |
878 | |
879 // Now remove the bogus extra edges used to keep things alive | |
880 if (can_reshape) { | |
881 phase->is_IterGVN()->remove_dead_node(hook); | |
882 } else { | |
883 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
884 } | |
885 return cmov2; | |
886 } | |
887 } | |
888 | |
889 // Fell thru, the unroll case is not appropriate. Transform the modulo | |
890 // into a long multiply/int multiply/subtract case | |
891 | |
892 // Cannot handle mod 0, and min_jint isn't handled by the transform | |
893 if( con == 0 || con == min_jint ) return NULL; | |
894 | |
895 // Get the absolute value of the constant; at this point, we can use this | |
896 jint pos_con = (con >= 0) ? con : -con; | |
897 | |
898 // integer Mod 1 is always 0 | |
899 if( pos_con == 1 ) return new (phase->C, 1) ConINode(TypeInt::ZERO); | |
900 | |
901 int log2_con = -1; | |
902 | |
903 // If this is a power of two, they maybe we can mask it | |
904 if( is_power_of_2(pos_con) ) { | |
905 log2_con = log2_intptr((intptr_t)pos_con); | |
906 | |
907 const Type *dt = phase->type(in(1)); | |
908 const TypeInt *dti = dt->isa_int(); | |
909 | |
910 // See if this can be masked, if the dividend is non-negative | |
911 if( dti && dti->_lo >= 0 ) | |
912 return ( new (phase->C, 3) AndINode( in(1), phase->intcon( pos_con-1 ) ) ); | |
913 } | |
914 | |
915 // Save in(1) so that it cannot be changed or deleted | |
916 hook->init_req(0, in(1)); | |
917 | |
918 // Divide using the transform from DivI to MulL | |
145 | 919 Node *result = transform_int_divide( phase, in(1), pos_con ); |
920 if (result != NULL) { | |
921 Node *divide = phase->transform(result); | |
0 | 922 |
145 | 923 // Re-multiply, using a shift if this is a power of two |
924 Node *mult = NULL; | |
0 | 925 |
145 | 926 if( log2_con >= 0 ) |
927 mult = phase->transform( new (phase->C, 3) LShiftINode( divide, phase->intcon( log2_con ) ) ); | |
928 else | |
929 mult = phase->transform( new (phase->C, 3) MulINode( divide, phase->intcon( pos_con ) ) ); | |
0 | 930 |
145 | 931 // Finally, subtract the multiplied divided value from the original |
932 result = new (phase->C, 3) SubINode( in(1), mult ); | |
933 } | |
0 | 934 |
935 // Now remove the bogus extra edges used to keep things alive | |
936 if (can_reshape) { | |
937 phase->is_IterGVN()->remove_dead_node(hook); | |
938 } else { | |
939 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
940 } | |
941 | |
942 // return the value | |
943 return result; | |
944 } | |
945 | |
946 //------------------------------Value------------------------------------------ | |
947 const Type *ModINode::Value( PhaseTransform *phase ) const { | |
948 // Either input is TOP ==> the result is TOP | |
949 const Type *t1 = phase->type( in(1) ); | |
950 const Type *t2 = phase->type( in(2) ); | |
951 if( t1 == Type::TOP ) return Type::TOP; | |
952 if( t2 == Type::TOP ) return Type::TOP; | |
953 | |
954 // We always generate the dynamic check for 0. | |
955 // 0 MOD X is 0 | |
956 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; | |
957 // X MOD X is 0 | |
958 if( phase->eqv( in(1), in(2) ) ) return TypeInt::ZERO; | |
959 | |
960 // Either input is BOTTOM ==> the result is the local BOTTOM | |
961 const Type *bot = bottom_type(); | |
962 if( (t1 == bot) || (t2 == bot) || | |
963 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
964 return bot; | |
965 | |
966 const TypeInt *i1 = t1->is_int(); | |
967 const TypeInt *i2 = t2->is_int(); | |
968 if( !i1->is_con() || !i2->is_con() ) { | |
969 if( i1->_lo >= 0 && i2->_lo >= 0 ) | |
970 return TypeInt::POS; | |
971 // If both numbers are not constants, we know little. | |
972 return TypeInt::INT; | |
973 } | |
974 // Mod by zero? Throw exception at runtime! | |
975 if( !i2->get_con() ) return TypeInt::POS; | |
976 | |
977 // We must be modulo'ing 2 float constants. | |
978 // Check for min_jint % '-1', result is defined to be '0'. | |
979 if( i1->get_con() == min_jint && i2->get_con() == -1 ) | |
980 return TypeInt::ZERO; | |
981 | |
982 return TypeInt::make( i1->get_con() % i2->get_con() ); | |
983 } | |
984 | |
985 | |
986 //============================================================================= | |
987 //------------------------------Idealize--------------------------------------- | |
988 Node *ModLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
989 // Check for dead control input | |
305 | 990 if( in(0) && remove_dead_region(phase, can_reshape) ) return this; |
991 // Don't bother trying to transform a dead node | |
992 if( in(0) && in(0)->is_top() ) return NULL; | |
0 | 993 |
994 // Get the modulus | |
995 const Type *t = phase->type( in(2) ); | |
996 if( t == Type::TOP ) return NULL; | |
145 | 997 const TypeLong *tl = t->is_long(); |
0 | 998 |
999 // Check for useless control input | |
1000 // Check for excluding mod-zero case | |
145 | 1001 if( in(0) && (tl->_hi < 0 || tl->_lo > 0) ) { |
0 | 1002 set_req(0, NULL); // Yank control input |
1003 return this; | |
1004 } | |
1005 | |
1006 // See if we are MOD'ing by 2^k or 2^k-1. | |
145 | 1007 if( !tl->is_con() ) return NULL; |
1008 jlong con = tl->get_con(); | |
1009 | |
1010 Node *hook = new (phase->C, 1) Node(1); | |
0 | 1011 |
1012 // Expand mod | |
145 | 1013 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
|
1014 uint k = exact_log2_long(con+1); // Extract k |
145 | 1015 |
0 | 1016 // Basic algorithm by David Detlefs. See fastmod_long.java for gory details. |
1017 // Used to help a popular random number generator which does a long-mod | |
1018 // of 2^31-1 and shows up in SpecJBB and SciMark. | |
1019 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*/}; | |
1020 int trip_count = 1; | |
1021 if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k]; | |
1022 | |
145 | 1023 // If the unroll factor is not too large, and if conditional moves are |
1024 // ok, then use this case | |
1025 if( trip_count <= 5 && ConditionalMoveLimit != 0 ) { | |
1026 Node *x = in(1); // Value being mod'd | |
1027 Node *divisor = in(2); // Also is mask | |
0 | 1028 |
145 | 1029 hook->init_req(0, x); // Add a use to x to prevent him from dying |
1030 // Generate code to reduce X rapidly to nearly 2^k-1. | |
1031 for( int i = 0; i < trip_count; i++ ) { | |
0 | 1032 Node *xl = phase->transform( new (phase->C, 3) AndLNode(x,divisor) ); |
1033 Node *xh = phase->transform( new (phase->C, 3) RShiftLNode(x,phase->intcon(k)) ); // Must be signed | |
1034 x = phase->transform( new (phase->C, 3) AddLNode(xh,xl) ); | |
1035 hook->set_req(0, x); // Add a use to x to prevent him from dying | |
145 | 1036 } |
1037 | |
1038 // Generate sign-fixup code. Was original value positive? | |
1039 // long hack_res = (i >= 0) ? divisor : CONST64(1); | |
1040 Node *cmp1 = phase->transform( new (phase->C, 3) CmpLNode( in(1), phase->longcon(0) ) ); | |
1041 Node *bol1 = phase->transform( new (phase->C, 2) BoolNode( cmp1, BoolTest::ge ) ); | |
1042 Node *cmov1= phase->transform( new (phase->C, 4) CMoveLNode(bol1, phase->longcon(1), divisor, TypeLong::LONG) ); | |
1043 // if( x >= hack_res ) x -= divisor; | |
1044 Node *sub = phase->transform( new (phase->C, 3) SubLNode( x, divisor ) ); | |
1045 Node *cmp2 = phase->transform( new (phase->C, 3) CmpLNode( x, cmov1 ) ); | |
1046 Node *bol2 = phase->transform( new (phase->C, 2) BoolNode( cmp2, BoolTest::ge ) ); | |
1047 // Convention is to not transform the return value of an Ideal | |
1048 // since Ideal is expected to return a modified 'this' or a new node. | |
1049 Node *cmov2= new (phase->C, 4) CMoveLNode(bol2, x, sub, TypeLong::LONG); | |
1050 // cmov2 is now the mod | |
1051 | |
1052 // Now remove the bogus extra edges used to keep things alive | |
1053 if (can_reshape) { | |
1054 phase->is_IterGVN()->remove_dead_node(hook); | |
1055 } else { | |
1056 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
1057 } | |
1058 return cmov2; | |
0 | 1059 } |
145 | 1060 } |
1061 | |
1062 // Fell thru, the unroll case is not appropriate. Transform the modulo | |
1063 // into a long multiply/int multiply/subtract case | |
1064 | |
1065 // Cannot handle mod 0, and min_jint isn't handled by the transform | |
1066 if( con == 0 || con == min_jlong ) return NULL; | |
1067 | |
1068 // Get the absolute value of the constant; at this point, we can use this | |
1069 jlong pos_con = (con >= 0) ? con : -con; | |
1070 | |
1071 // integer Mod 1 is always 0 | |
1072 if( pos_con == 1 ) return new (phase->C, 1) ConLNode(TypeLong::ZERO); | |
1073 | |
1074 int log2_con = -1; | |
1075 | |
605 | 1076 // If this is a power of two, then maybe we can mask it |
145 | 1077 if( is_power_of_2_long(pos_con) ) { |
1078 log2_con = log2_long(pos_con); | |
1079 | |
1080 const Type *dt = phase->type(in(1)); | |
1081 const TypeLong *dtl = dt->isa_long(); | |
1082 | |
1083 // See if this can be masked, if the dividend is non-negative | |
1084 if( dtl && dtl->_lo >= 0 ) | |
1085 return ( new (phase->C, 3) AndLNode( in(1), phase->longcon( pos_con-1 ) ) ); | |
1086 } | |
0 | 1087 |
145 | 1088 // Save in(1) so that it cannot be changed or deleted |
1089 hook->init_req(0, in(1)); | |
1090 | |
1091 // Divide using the transform from DivI to MulL | |
1092 Node *result = transform_long_divide( phase, in(1), pos_con ); | |
1093 if (result != NULL) { | |
1094 Node *divide = phase->transform(result); | |
1095 | |
1096 // Re-multiply, using a shift if this is a power of two | |
1097 Node *mult = NULL; | |
1098 | |
1099 if( log2_con >= 0 ) | |
1100 mult = phase->transform( new (phase->C, 3) LShiftLNode( divide, phase->intcon( log2_con ) ) ); | |
1101 else | |
1102 mult = phase->transform( new (phase->C, 3) MulLNode( divide, phase->longcon( pos_con ) ) ); | |
1103 | |
1104 // Finally, subtract the multiplied divided value from the original | |
1105 result = new (phase->C, 3) SubLNode( in(1), mult ); | |
0 | 1106 } |
145 | 1107 |
1108 // Now remove the bogus extra edges used to keep things alive | |
1109 if (can_reshape) { | |
1110 phase->is_IterGVN()->remove_dead_node(hook); | |
1111 } else { | |
1112 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
1113 } | |
1114 | |
1115 // return the value | |
1116 return result; | |
0 | 1117 } |
1118 | |
1119 //------------------------------Value------------------------------------------ | |
1120 const Type *ModLNode::Value( PhaseTransform *phase ) const { | |
1121 // Either input is TOP ==> the result is TOP | |
1122 const Type *t1 = phase->type( in(1) ); | |
1123 const Type *t2 = phase->type( in(2) ); | |
1124 if( t1 == Type::TOP ) return Type::TOP; | |
1125 if( t2 == Type::TOP ) return Type::TOP; | |
1126 | |
1127 // We always generate the dynamic check for 0. | |
1128 // 0 MOD X is 0 | |
1129 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; | |
1130 // X MOD X is 0 | |
1131 if( phase->eqv( in(1), in(2) ) ) return TypeLong::ZERO; | |
1132 | |
1133 // Either input is BOTTOM ==> the result is the local BOTTOM | |
1134 const Type *bot = bottom_type(); | |
1135 if( (t1 == bot) || (t2 == bot) || | |
1136 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
1137 return bot; | |
1138 | |
1139 const TypeLong *i1 = t1->is_long(); | |
1140 const TypeLong *i2 = t2->is_long(); | |
1141 if( !i1->is_con() || !i2->is_con() ) { | |
1142 if( i1->_lo >= CONST64(0) && i2->_lo >= CONST64(0) ) | |
1143 return TypeLong::POS; | |
1144 // If both numbers are not constants, we know little. | |
1145 return TypeLong::LONG; | |
1146 } | |
1147 // Mod by zero? Throw exception at runtime! | |
1148 if( !i2->get_con() ) return TypeLong::POS; | |
1149 | |
1150 // We must be modulo'ing 2 float constants. | |
1151 // Check for min_jint % '-1', result is defined to be '0'. | |
1152 if( i1->get_con() == min_jlong && i2->get_con() == -1 ) | |
1153 return TypeLong::ZERO; | |
1154 | |
1155 return TypeLong::make( i1->get_con() % i2->get_con() ); | |
1156 } | |
1157 | |
1158 | |
1159 //============================================================================= | |
1160 //------------------------------Value------------------------------------------ | |
1161 const Type *ModFNode::Value( PhaseTransform *phase ) const { | |
1162 // Either input is TOP ==> the result is TOP | |
1163 const Type *t1 = phase->type( in(1) ); | |
1164 const Type *t2 = phase->type( in(2) ); | |
1165 if( t1 == Type::TOP ) return Type::TOP; | |
1166 if( t2 == Type::TOP ) return Type::TOP; | |
1167 | |
1168 // Either input is BOTTOM ==> the result is the local BOTTOM | |
1169 const Type *bot = bottom_type(); | |
1170 if( (t1 == bot) || (t2 == bot) || | |
1171 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
1172 return bot; | |
1173 | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1174 // 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
|
1175 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
|
1176 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
|
1177 } |
0 | 1178 |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1179 float f1 = t1->getf(); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1180 float f2 = t2->getf(); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1181 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
|
1182 jint x2 = jint_cast(f2); |
0 | 1183 |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1184 // 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
|
1185 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
|
1186 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
|
1187 |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1188 // 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
|
1189 if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jint) |
0 | 1190 return Type::FLOAT; |
1191 | |
1192 // We must be modulo'ing 2 float constants. | |
1193 // 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
|
1194 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
|
1195 if ((x1 ^ xr) < 0) { |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1196 xr ^= min_jint; |
0 | 1197 } |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1198 |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1199 return TypeF::make(jfloat_cast(xr)); |
0 | 1200 } |
1201 | |
1202 | |
1203 //============================================================================= | |
1204 //------------------------------Value------------------------------------------ | |
1205 const Type *ModDNode::Value( PhaseTransform *phase ) const { | |
1206 // Either input is TOP ==> the result is TOP | |
1207 const Type *t1 = phase->type( in(1) ); | |
1208 const Type *t2 = phase->type( in(2) ); | |
1209 if( t1 == Type::TOP ) return Type::TOP; | |
1210 if( t2 == Type::TOP ) return Type::TOP; | |
1211 | |
1212 // Either input is BOTTOM ==> the result is the local BOTTOM | |
1213 const Type *bot = bottom_type(); | |
1214 if( (t1 == bot) || (t2 == bot) || | |
1215 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
1216 return bot; | |
1217 | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1218 // 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
|
1219 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
|
1220 return Type::DOUBLE; // note: x%x can be either NaN or 0 |
0 | 1221 } |
1222 | |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1223 double f1 = t1->getd(); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1224 double f2 = t2->getd(); |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1225 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
|
1226 jlong x2 = jlong_cast(f2); |
0 | 1227 |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1228 // 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
|
1229 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
|
1230 if (g_isnan(f2)) return t2; |
0 | 1231 |
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1232 // 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
|
1233 if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jlong) |
0 | 1234 return Type::DOUBLE; |
1235 | |
1236 // 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
|
1237 // 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
|
1238 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
|
1239 if ((x1 ^ xr) < 0) { |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1240 xr ^= min_jlong; |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1241 } |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1242 |
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1243 return TypeD::make(jdouble_cast(xr)); |
0 | 1244 } |
1245 | |
1246 //============================================================================= | |
1247 | |
1248 DivModNode::DivModNode( Node *c, Node *dividend, Node *divisor ) : MultiNode(3) { | |
1249 init_req(0, c); | |
1250 init_req(1, dividend); | |
1251 init_req(2, divisor); | |
1252 } | |
1253 | |
1254 //------------------------------make------------------------------------------ | |
1255 DivModINode* DivModINode::make(Compile* C, Node* div_or_mod) { | |
1256 Node* n = div_or_mod; | |
1257 assert(n->Opcode() == Op_DivI || n->Opcode() == Op_ModI, | |
1258 "only div or mod input pattern accepted"); | |
1259 | |
1260 DivModINode* divmod = new (C, 3) DivModINode(n->in(0), n->in(1), n->in(2)); | |
1261 Node* dproj = new (C, 1) ProjNode(divmod, DivModNode::div_proj_num); | |
1262 Node* mproj = new (C, 1) ProjNode(divmod, DivModNode::mod_proj_num); | |
1263 return divmod; | |
1264 } | |
1265 | |
1266 //------------------------------make------------------------------------------ | |
1267 DivModLNode* DivModLNode::make(Compile* C, Node* div_or_mod) { | |
1268 Node* n = div_or_mod; | |
1269 assert(n->Opcode() == Op_DivL || n->Opcode() == Op_ModL, | |
1270 "only div or mod input pattern accepted"); | |
1271 | |
1272 DivModLNode* divmod = new (C, 3) DivModLNode(n->in(0), n->in(1), n->in(2)); | |
1273 Node* dproj = new (C, 1) ProjNode(divmod, DivModNode::div_proj_num); | |
1274 Node* mproj = new (C, 1) ProjNode(divmod, DivModNode::mod_proj_num); | |
1275 return divmod; | |
1276 } | |
1277 | |
1278 //------------------------------match------------------------------------------ | |
1279 // return result(s) along with their RegMask info | |
1280 Node *DivModINode::match( const ProjNode *proj, const Matcher *match ) { | |
1281 uint ideal_reg = proj->ideal_reg(); | |
1282 RegMask rm; | |
1283 if (proj->_con == div_proj_num) { | |
1284 rm = match->divI_proj_mask(); | |
1285 } else { | |
1286 assert(proj->_con == mod_proj_num, "must be div or mod projection"); | |
1287 rm = match->modI_proj_mask(); | |
1288 } | |
1289 return new (match->C, 1)MachProjNode(this, proj->_con, rm, ideal_reg); | |
1290 } | |
1291 | |
1292 | |
1293 //------------------------------match------------------------------------------ | |
1294 // return result(s) along with their RegMask info | |
1295 Node *DivModLNode::match( const ProjNode *proj, const Matcher *match ) { | |
1296 uint ideal_reg = proj->ideal_reg(); | |
1297 RegMask rm; | |
1298 if (proj->_con == div_proj_num) { | |
1299 rm = match->divL_proj_mask(); | |
1300 } else { | |
1301 assert(proj->_con == mod_proj_num, "must be div or mod projection"); | |
1302 rm = match->modL_proj_mask(); | |
1303 } | |
1304 return new (match->C, 1)MachProjNode(this, proj->_con, rm, ideal_reg); | |
1305 } |