Mercurial > hg > truffle
annotate src/share/vm/opto/divnode.cpp @ 756:cecd04fc6f93
6837011: SIGSEGV in PhaseIdealLoop in 32bit jvm
Summary: The CR's test crashes with SIGSEGV when running with "-server -Xcomp" using using 32bit jvm.
Reviewed-by: kvn, never, rasbold
| author | twisti |
|---|---|
| date | Wed, 06 May 2009 12:04:42 -0700 |
| parents | 98cb887364d3 |
| children | 174ade00803b |
| 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. |
| 756 | 301 // 6837011: We need to transform w2 before w1 because the |
| 302 // transformation of w1 could return t. | |
| 303 w2 = phase->transform(w2); | |
|
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
304 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
|
305 |
|
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
306 // w1 = u0*v1 + w1; |
|
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
307 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
|
308 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
|
309 |
|
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
310 // return u1*v1 + w2 + (w1 >> 32); |
|
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
311 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
|
312 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
|
313 Node* temp2 = phase->transform(new (phase->C, 3) RShiftLNode(w1, phase->intcon(N / 2))); |
| 145 | 314 |
|
567
bbef4344adb2
6800154: Add comments to long_by_long_mulhi() for better understandability
twisti
parents:
404
diff
changeset
|
315 return new (phase->C, 3) AddLNode(temp1, temp2); |
| 145 | 316 } |
| 317 | |
| 318 | |
| 319 //--------------------------transform_long_divide------------------------------ | |
| 320 // Convert a division by constant divisor into an alternate Ideal graph. | |
| 321 // Return NULL if no transformation occurs. | |
| 322 static Node *transform_long_divide( PhaseGVN *phase, Node *dividend, jlong divisor ) { | |
| 323 // Check for invalid divisors | |
| 324 assert( divisor != 0L && divisor != min_jlong, | |
| 325 "bad divisor for transforming to long multiply" ); | |
| 326 | |
| 327 bool d_pos = divisor >= 0; | |
| 328 jlong d = d_pos ? divisor : -divisor; | |
| 329 const int N = 64; | |
| 330 | |
| 331 // Result | |
| 332 Node *q = NULL; | |
| 333 | |
| 334 if (d == 1) { | |
| 335 // division by +/- 1 | |
| 336 if (!d_pos) { | |
| 337 // Just negate the value | |
| 338 q = new (phase->C, 3) SubLNode(phase->longcon(0), dividend); | |
| 339 } | |
| 340 } else if ( is_power_of_2_long(d) ) { | |
| 341 | |
| 342 // division by +/- a power of 2 | |
| 343 | |
| 344 // See if we can simply do a shift without rounding | |
| 345 bool needs_rounding = true; | |
| 346 const Type *dt = phase->type(dividend); | |
| 347 const TypeLong *dtl = dt->isa_long(); | |
| 0 | 348 |
| 145 | 349 if (dtl && dtl->_lo > 0) { |
| 350 // we don't need to round a positive dividend | |
| 351 needs_rounding = false; | |
| 352 } else if( dividend->Opcode() == Op_AndL ) { | |
| 353 // An AND mask of sufficient size clears the low bits and | |
| 354 // I can avoid rounding. | |
|
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
355 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
|
356 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
|
357 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
|
358 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
|
359 dividend = dividend->in(1); |
|
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
360 needs_rounding = false; |
|
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
305
diff
changeset
|
361 } |
| 145 | 362 } |
| 363 } | |
| 364 | |
| 365 // Add rounding to the shift to handle the sign bit | |
| 366 int l = log2_long(d-1)+1; | |
| 367 if (needs_rounding) { | |
| 368 // Divide-by-power-of-2 can be made into a shift, but you have to do | |
| 369 // more math for the rounding. You need to add 0 for positive | |
| 370 // numbers, and "i-1" for negative numbers. Example: i=4, so the | |
| 371 // shift is by 2. You need to add 3 to negative dividends and 0 to | |
| 372 // positive ones. So (-7+3)>>2 becomes -1, (-4+3)>>2 becomes -1, | |
| 373 // (-2+3)>>2 becomes 0, etc. | |
| 374 | |
| 375 // Compute 0 or -1, based on sign bit | |
| 376 Node *sign = phase->transform(new (phase->C, 3) RShiftLNode(dividend, phase->intcon(N - 1))); | |
| 377 // Mask sign bit to the low sign bits | |
| 378 Node *round = phase->transform(new (phase->C, 3) URShiftLNode(sign, phase->intcon(N - l))); | |
| 379 // Round up before shifting | |
| 380 dividend = phase->transform(new (phase->C, 3) AddLNode(dividend, round)); | |
| 381 } | |
| 382 | |
| 383 // Shift for division | |
| 384 q = new (phase->C, 3) RShiftLNode(dividend, phase->intcon(l)); | |
| 385 | |
| 386 if (!d_pos) { | |
| 387 q = new (phase->C, 3) SubLNode(phase->longcon(0), phase->transform(q)); | |
| 388 } | |
| 389 } else { | |
| 390 // Attempt the jlong constant divide -> multiply transform found in | |
| 391 // "Division by Invariant Integers using Multiplication" | |
| 392 // by Granlund and Montgomery | |
| 393 // See also "Hacker's Delight", chapter 10 by Warren. | |
| 394 | |
| 395 jlong magic_const; | |
| 396 jint shift_const; | |
| 397 if (magic_long_divide_constants(d, magic_const, shift_const)) { | |
| 398 // Compute the high half of the dividend x magic multiplication | |
| 399 Node *mul_hi = phase->transform(long_by_long_mulhi(phase, dividend, magic_const)); | |
| 400 | |
| 401 // The high half of the 128-bit multiply is computed. | |
| 402 if (magic_const < 0) { | |
| 403 // The magic multiplier is too large for a 64 bit constant. We've adjusted | |
| 404 // it down by 2^64, but have to add 1 dividend back in after the multiplication. | |
| 405 // This handles the "overflow" case described by Granlund and Montgomery. | |
| 406 mul_hi = phase->transform(new (phase->C, 3) AddLNode(dividend, mul_hi)); | |
| 407 } | |
| 408 | |
| 409 // Shift over the (adjusted) mulhi | |
| 410 if (shift_const != 0) { | |
| 411 mul_hi = phase->transform(new (phase->C, 3) RShiftLNode(mul_hi, phase->intcon(shift_const))); | |
| 412 } | |
| 413 | |
| 414 // Get a 0 or -1 from the sign of the dividend. | |
| 415 Node *addend0 = mul_hi; | |
| 416 Node *addend1 = phase->transform(new (phase->C, 3) RShiftLNode(dividend, phase->intcon(N-1))); | |
| 417 | |
| 418 // If the divisor is negative, swap the order of the input addends; | |
| 419 // this has the effect of negating the quotient. | |
| 420 if (!d_pos) { | |
| 421 Node *temp = addend0; addend0 = addend1; addend1 = temp; | |
| 422 } | |
| 423 | |
| 424 // Adjust the final quotient by subtracting -1 (adding 1) | |
| 425 // from the mul_hi. | |
| 426 q = new (phase->C, 3) SubLNode(addend0, addend1); | |
| 427 } | |
| 0 | 428 } |
| 429 | |
| 145 | 430 return q; |
| 0 | 431 } |
| 432 | |
| 433 //============================================================================= | |
| 434 //------------------------------Identity--------------------------------------- | |
| 435 // If the divisor is 1, we are an identity on the dividend. | |
| 436 Node *DivINode::Identity( PhaseTransform *phase ) { | |
| 437 return (phase->type( in(2) )->higher_equal(TypeInt::ONE)) ? in(1) : this; | |
| 438 } | |
| 439 | |
| 440 //------------------------------Idealize--------------------------------------- | |
| 441 // Divides can be changed to multiplies and/or shifts | |
| 442 Node *DivINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
| 443 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
| 305 | 444 // Don't bother trying to transform a dead node |
| 445 if( in(0) && in(0)->is_top() ) return NULL; | |
| 0 | 446 |
| 447 const Type *t = phase->type( in(2) ); | |
| 448 if( t == TypeInt::ONE ) // Identity? | |
| 449 return NULL; // Skip it | |
| 450 | |
| 451 const TypeInt *ti = t->isa_int(); | |
| 452 if( !ti ) return NULL; | |
| 453 if( !ti->is_con() ) return NULL; | |
| 145 | 454 jint i = ti->get_con(); // Get divisor |
| 0 | 455 |
| 456 if (i == 0) return NULL; // Dividing by zero constant does not idealize | |
| 457 | |
| 458 set_req(0,NULL); // Dividing by a not-zero constant; no faulting | |
| 459 | |
| 460 // Dividing by MININT does not optimize as a power-of-2 shift. | |
| 461 if( i == min_jint ) return NULL; | |
| 462 | |
| 145 | 463 return transform_int_divide( phase, in(1), i ); |
| 0 | 464 } |
| 465 | |
| 466 //------------------------------Value------------------------------------------ | |
| 467 // A DivINode divides its inputs. The third input is a Control input, used to | |
| 468 // prevent hoisting the divide above an unsafe test. | |
| 469 const Type *DivINode::Value( PhaseTransform *phase ) const { | |
| 470 // Either input is TOP ==> the result is TOP | |
| 471 const Type *t1 = phase->type( in(1) ); | |
| 472 const Type *t2 = phase->type( in(2) ); | |
| 473 if( t1 == Type::TOP ) return Type::TOP; | |
| 474 if( t2 == Type::TOP ) return Type::TOP; | |
| 475 | |
| 476 // x/x == 1 since we always generate the dynamic divisor check for 0. | |
| 477 if( phase->eqv( in(1), in(2) ) ) | |
| 478 return TypeInt::ONE; | |
| 479 | |
| 480 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 481 const Type *bot = bottom_type(); | |
| 482 if( (t1 == bot) || (t2 == bot) || | |
| 483 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 484 return bot; | |
| 485 | |
| 486 // Divide the two numbers. We approximate. | |
| 487 // If divisor is a constant and not zero | |
| 488 const TypeInt *i1 = t1->is_int(); | |
| 489 const TypeInt *i2 = t2->is_int(); | |
| 490 int widen = MAX2(i1->_widen, i2->_widen); | |
| 491 | |
| 492 if( i2->is_con() && i2->get_con() != 0 ) { | |
| 493 int32 d = i2->get_con(); // Divisor | |
| 494 jint lo, hi; | |
| 495 if( d >= 0 ) { | |
| 496 lo = i1->_lo/d; | |
| 497 hi = i1->_hi/d; | |
| 498 } else { | |
| 499 if( d == -1 && i1->_lo == min_jint ) { | |
| 500 // 'min_jint/-1' throws arithmetic exception during compilation | |
| 501 lo = min_jint; | |
| 502 // do not support holes, 'hi' must go to either min_jint or max_jint: | |
| 503 // [min_jint, -10]/[-1,-1] ==> [min_jint] UNION [10,max_jint] | |
| 504 hi = i1->_hi == min_jint ? min_jint : max_jint; | |
| 505 } else { | |
| 506 lo = i1->_hi/d; | |
| 507 hi = i1->_lo/d; | |
| 508 } | |
| 509 } | |
| 510 return TypeInt::make(lo, hi, widen); | |
| 511 } | |
| 512 | |
| 513 // If the dividend is a constant | |
| 514 if( i1->is_con() ) { | |
| 515 int32 d = i1->get_con(); | |
| 516 if( d < 0 ) { | |
| 517 if( d == min_jint ) { | |
| 518 // (-min_jint) == min_jint == (min_jint / -1) | |
| 519 return TypeInt::make(min_jint, max_jint/2 + 1, widen); | |
| 520 } else { | |
| 521 return TypeInt::make(d, -d, widen); | |
| 522 } | |
| 523 } | |
| 524 return TypeInt::make(-d, d, widen); | |
| 525 } | |
| 526 | |
| 527 // Otherwise we give up all hope | |
| 528 return TypeInt::INT; | |
| 529 } | |
| 530 | |
| 531 | |
| 532 //============================================================================= | |
| 533 //------------------------------Identity--------------------------------------- | |
| 534 // If the divisor is 1, we are an identity on the dividend. | |
| 535 Node *DivLNode::Identity( PhaseTransform *phase ) { | |
| 536 return (phase->type( in(2) )->higher_equal(TypeLong::ONE)) ? in(1) : this; | |
| 537 } | |
| 538 | |
| 539 //------------------------------Idealize--------------------------------------- | |
| 540 // Dividing by a power of 2 is a shift. | |
| 541 Node *DivLNode::Ideal( PhaseGVN *phase, bool can_reshape) { | |
| 542 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
| 305 | 543 // Don't bother trying to transform a dead node |
| 544 if( in(0) && in(0)->is_top() ) return NULL; | |
| 0 | 545 |
| 546 const Type *t = phase->type( in(2) ); | |
| 145 | 547 if( t == TypeLong::ONE ) // Identity? |
| 0 | 548 return NULL; // Skip it |
| 549 | |
| 145 | 550 const TypeLong *tl = t->isa_long(); |
| 551 if( !tl ) return NULL; | |
| 552 if( !tl->is_con() ) return NULL; | |
| 553 jlong l = tl->get_con(); // Get divisor | |
| 554 | |
| 555 if (l == 0) return NULL; // Dividing by zero constant does not idealize | |
| 556 | |
| 557 set_req(0,NULL); // Dividing by a not-zero constant; no faulting | |
| 0 | 558 |
| 559 // Dividing by MININT does not optimize as a power-of-2 shift. | |
| 145 | 560 if( l == min_jlong ) return NULL; |
| 0 | 561 |
| 145 | 562 return transform_long_divide( phase, in(1), l ); |
| 0 | 563 } |
| 564 | |
| 565 //------------------------------Value------------------------------------------ | |
| 566 // A DivLNode divides its inputs. The third input is a Control input, used to | |
| 567 // prevent hoisting the divide above an unsafe test. | |
| 568 const Type *DivLNode::Value( PhaseTransform *phase ) const { | |
| 569 // Either input is TOP ==> the result is TOP | |
| 570 const Type *t1 = phase->type( in(1) ); | |
| 571 const Type *t2 = phase->type( in(2) ); | |
| 572 if( t1 == Type::TOP ) return Type::TOP; | |
| 573 if( t2 == Type::TOP ) return Type::TOP; | |
| 574 | |
| 575 // x/x == 1 since we always generate the dynamic divisor check for 0. | |
| 576 if( phase->eqv( in(1), in(2) ) ) | |
| 577 return TypeLong::ONE; | |
| 578 | |
| 579 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 580 const Type *bot = bottom_type(); | |
| 581 if( (t1 == bot) || (t2 == bot) || | |
| 582 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 583 return bot; | |
| 584 | |
| 585 // Divide the two numbers. We approximate. | |
| 586 // If divisor is a constant and not zero | |
| 587 const TypeLong *i1 = t1->is_long(); | |
| 588 const TypeLong *i2 = t2->is_long(); | |
| 589 int widen = MAX2(i1->_widen, i2->_widen); | |
| 590 | |
| 591 if( i2->is_con() && i2->get_con() != 0 ) { | |
| 592 jlong d = i2->get_con(); // Divisor | |
| 593 jlong lo, hi; | |
| 594 if( d >= 0 ) { | |
| 595 lo = i1->_lo/d; | |
| 596 hi = i1->_hi/d; | |
| 597 } else { | |
| 598 if( d == CONST64(-1) && i1->_lo == min_jlong ) { | |
| 599 // 'min_jlong/-1' throws arithmetic exception during compilation | |
| 600 lo = min_jlong; | |
| 601 // do not support holes, 'hi' must go to either min_jlong or max_jlong: | |
| 602 // [min_jlong, -10]/[-1,-1] ==> [min_jlong] UNION [10,max_jlong] | |
| 603 hi = i1->_hi == min_jlong ? min_jlong : max_jlong; | |
| 604 } else { | |
| 605 lo = i1->_hi/d; | |
| 606 hi = i1->_lo/d; | |
| 607 } | |
| 608 } | |
| 609 return TypeLong::make(lo, hi, widen); | |
| 610 } | |
| 611 | |
| 612 // If the dividend is a constant | |
| 613 if( i1->is_con() ) { | |
| 614 jlong d = i1->get_con(); | |
| 615 if( d < 0 ) { | |
| 616 if( d == min_jlong ) { | |
| 617 // (-min_jlong) == min_jlong == (min_jlong / -1) | |
| 618 return TypeLong::make(min_jlong, max_jlong/2 + 1, widen); | |
| 619 } else { | |
| 620 return TypeLong::make(d, -d, widen); | |
| 621 } | |
| 622 } | |
| 623 return TypeLong::make(-d, d, widen); | |
| 624 } | |
| 625 | |
| 626 // Otherwise we give up all hope | |
| 627 return TypeLong::LONG; | |
| 628 } | |
| 629 | |
| 630 | |
| 631 //============================================================================= | |
| 632 //------------------------------Value------------------------------------------ | |
| 633 // An DivFNode divides its inputs. The third input is a Control input, used to | |
| 634 // prevent hoisting the divide above an unsafe test. | |
| 635 const Type *DivFNode::Value( PhaseTransform *phase ) const { | |
| 636 // Either input is TOP ==> the result is TOP | |
| 637 const Type *t1 = phase->type( in(1) ); | |
| 638 const Type *t2 = phase->type( in(2) ); | |
| 639 if( t1 == Type::TOP ) return Type::TOP; | |
| 640 if( t2 == Type::TOP ) return Type::TOP; | |
| 641 | |
| 642 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 643 const Type *bot = bottom_type(); | |
| 644 if( (t1 == bot) || (t2 == bot) || | |
| 645 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 646 return bot; | |
| 647 | |
| 648 // x/x == 1, we ignore 0/0. | |
| 649 // 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
|
650 // Does not work for variables because of NaN's |
| 0 | 651 if( phase->eqv( in(1), in(2) ) && t1->base() == Type::FloatCon) |
| 652 if (!g_isnan(t1->getf()) && g_isfinite(t1->getf()) && t1->getf() != 0.0) // could be negative ZERO or NaN | |
| 653 return TypeF::ONE; | |
| 654 | |
| 655 if( t2 == TypeF::ONE ) | |
| 656 return t1; | |
| 657 | |
| 658 // If divisor is a constant and not zero, divide them numbers | |
| 659 if( t1->base() == Type::FloatCon && | |
| 660 t2->base() == Type::FloatCon && | |
| 661 t2->getf() != 0.0 ) // could be negative zero | |
| 662 return TypeF::make( t1->getf()/t2->getf() ); | |
| 663 | |
| 664 // If the dividend is a constant zero | |
| 665 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
| 666 // Test TypeF::ZERO is not sufficient as it could be negative zero | |
| 667 | |
| 668 if( t1 == TypeF::ZERO && !g_isnan(t2->getf()) && t2->getf() != 0.0 ) | |
| 669 return TypeF::ZERO; | |
| 670 | |
| 671 // Otherwise we give up all hope | |
| 672 return Type::FLOAT; | |
| 673 } | |
| 674 | |
| 675 //------------------------------isA_Copy--------------------------------------- | |
| 676 // Dividing by self is 1. | |
| 677 // If the divisor is 1, we are an identity on the dividend. | |
| 678 Node *DivFNode::Identity( PhaseTransform *phase ) { | |
| 679 return (phase->type( in(2) ) == TypeF::ONE) ? in(1) : this; | |
| 680 } | |
| 681 | |
| 682 | |
| 683 //------------------------------Idealize--------------------------------------- | |
| 684 Node *DivFNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
| 685 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
| 305 | 686 // Don't bother trying to transform a dead node |
| 687 if( in(0) && in(0)->is_top() ) return NULL; | |
| 0 | 688 |
| 689 const Type *t2 = phase->type( in(2) ); | |
| 690 if( t2 == TypeF::ONE ) // Identity? | |
| 691 return NULL; // Skip it | |
| 692 | |
| 693 const TypeF *tf = t2->isa_float_constant(); | |
| 694 if( !tf ) return NULL; | |
| 695 if( tf->base() != Type::FloatCon ) return NULL; | |
| 696 | |
| 697 // Check for out of range values | |
| 698 if( tf->is_nan() || !tf->is_finite() ) return NULL; | |
| 699 | |
| 700 // Get the value | |
| 701 float f = tf->getf(); | |
| 702 int exp; | |
| 703 | |
| 704 // Only for special case of dividing by a power of 2 | |
| 705 if( frexp((double)f, &exp) != 0.5 ) return NULL; | |
| 706 | |
| 707 // Limit the range of acceptable exponents | |
| 708 if( exp < -126 || exp > 126 ) return NULL; | |
| 709 | |
| 710 // Compute the reciprocal | |
| 711 float reciprocal = ((float)1.0) / f; | |
| 712 | |
| 713 assert( frexp((double)reciprocal, &exp) == 0.5, "reciprocal should be power of 2" ); | |
| 714 | |
| 715 // return multiplication by the reciprocal | |
| 716 return (new (phase->C, 3) MulFNode(in(1), phase->makecon(TypeF::make(reciprocal)))); | |
| 717 } | |
| 718 | |
| 719 //============================================================================= | |
| 720 //------------------------------Value------------------------------------------ | |
| 721 // 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
|
722 // prevent hoisting the divide above an unsafe test. |
| 0 | 723 const Type *DivDNode::Value( PhaseTransform *phase ) const { |
| 724 // Either input is TOP ==> the result is TOP | |
| 725 const Type *t1 = phase->type( in(1) ); | |
| 726 const Type *t2 = phase->type( in(2) ); | |
| 727 if( t1 == Type::TOP ) return Type::TOP; | |
| 728 if( t2 == Type::TOP ) return Type::TOP; | |
| 729 | |
| 730 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 731 const Type *bot = bottom_type(); | |
| 732 if( (t1 == bot) || (t2 == bot) || | |
| 733 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 734 return bot; | |
| 735 | |
| 736 // x/x == 1, we ignore 0/0. | |
| 737 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
| 738 // Does not work for variables because of NaN's | |
| 739 if( phase->eqv( in(1), in(2) ) && t1->base() == Type::DoubleCon) | |
| 740 if (!g_isnan(t1->getd()) && g_isfinite(t1->getd()) && t1->getd() != 0.0) // could be negative ZERO or NaN | |
| 741 return TypeD::ONE; | |
| 742 | |
| 743 if( t2 == TypeD::ONE ) | |
| 744 return t1; | |
| 745 | |
|
404
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
746 #if defined(IA32) |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
747 if (!phase->C->method()->is_strict()) |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
748 // 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
|
749 // division with round-to-zero on this platform. |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
750 #endif |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
751 { |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
752 // 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
|
753 if( t1->base() == Type::DoubleCon && |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
754 t2->base() == Type::DoubleCon && |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
755 t2->getd() != 0.0 ) // could be negative zero |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
756 return TypeD::make( t1->getd()/t2->getd() ); |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
757 } |
| 0 | 758 |
| 759 // If the dividend is a constant zero | |
| 760 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
| 761 // Test TypeF::ZERO is not sufficient as it could be negative zero | |
| 762 if( t1 == TypeD::ZERO && !g_isnan(t2->getd()) && t2->getd() != 0.0 ) | |
| 763 return TypeD::ZERO; | |
| 764 | |
| 765 // Otherwise we give up all hope | |
| 766 return Type::DOUBLE; | |
| 767 } | |
| 768 | |
| 769 | |
| 770 //------------------------------isA_Copy--------------------------------------- | |
| 771 // Dividing by self is 1. | |
| 772 // If the divisor is 1, we are an identity on the dividend. | |
| 773 Node *DivDNode::Identity( PhaseTransform *phase ) { | |
| 774 return (phase->type( in(2) ) == TypeD::ONE) ? in(1) : this; | |
| 775 } | |
| 776 | |
| 777 //------------------------------Idealize--------------------------------------- | |
| 778 Node *DivDNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
| 779 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
| 305 | 780 // Don't bother trying to transform a dead node |
| 781 if( in(0) && in(0)->is_top() ) return NULL; | |
| 0 | 782 |
| 783 const Type *t2 = phase->type( in(2) ); | |
| 784 if( t2 == TypeD::ONE ) // Identity? | |
| 785 return NULL; // Skip it | |
| 786 | |
| 787 const TypeD *td = t2->isa_double_constant(); | |
| 788 if( !td ) return NULL; | |
| 789 if( td->base() != Type::DoubleCon ) return NULL; | |
| 790 | |
| 791 // Check for out of range values | |
| 792 if( td->is_nan() || !td->is_finite() ) return NULL; | |
| 793 | |
| 794 // Get the value | |
| 795 double d = td->getd(); | |
| 796 int exp; | |
| 797 | |
| 798 // Only for special case of dividing by a power of 2 | |
| 799 if( frexp(d, &exp) != 0.5 ) return NULL; | |
| 800 | |
| 801 // Limit the range of acceptable exponents | |
| 802 if( exp < -1021 || exp > 1022 ) return NULL; | |
| 803 | |
| 804 // Compute the reciprocal | |
| 805 double reciprocal = 1.0 / d; | |
| 806 | |
| 807 assert( frexp(reciprocal, &exp) == 0.5, "reciprocal should be power of 2" ); | |
| 808 | |
| 809 // return multiplication by the reciprocal | |
| 810 return (new (phase->C, 3) MulDNode(in(1), phase->makecon(TypeD::make(reciprocal)))); | |
| 811 } | |
| 812 | |
| 813 //============================================================================= | |
| 814 //------------------------------Idealize--------------------------------------- | |
| 815 Node *ModINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
| 816 // Check for dead control input | |
| 305 | 817 if( in(0) && remove_dead_region(phase, can_reshape) ) return this; |
| 818 // Don't bother trying to transform a dead node | |
| 819 if( in(0) && in(0)->is_top() ) return NULL; | |
| 0 | 820 |
| 821 // Get the modulus | |
| 822 const Type *t = phase->type( in(2) ); | |
| 823 if( t == Type::TOP ) return NULL; | |
| 824 const TypeInt *ti = t->is_int(); | |
| 825 | |
| 826 // Check for useless control input | |
| 827 // Check for excluding mod-zero case | |
| 828 if( in(0) && (ti->_hi < 0 || ti->_lo > 0) ) { | |
| 829 set_req(0, NULL); // Yank control input | |
| 830 return this; | |
| 831 } | |
| 832 | |
| 833 // See if we are MOD'ing by 2^k or 2^k-1. | |
| 834 if( !ti->is_con() ) return NULL; | |
| 835 jint con = ti->get_con(); | |
| 836 | |
| 837 Node *hook = new (phase->C, 1) Node(1); | |
| 838 | |
| 839 // First, special check for modulo 2^k-1 | |
| 840 if( con >= 0 && con < max_jint && is_power_of_2(con+1) ) { | |
| 841 uint k = exact_log2(con+1); // Extract k | |
| 842 | |
| 843 // Basic algorithm by David Detlefs. See fastmod_int.java for gory details. | |
| 844 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*/}; | |
| 845 int trip_count = 1; | |
| 846 if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k]; | |
| 847 | |
| 848 // If the unroll factor is not too large, and if conditional moves are | |
| 849 // ok, then use this case | |
| 850 if( trip_count <= 5 && ConditionalMoveLimit != 0 ) { | |
| 851 Node *x = in(1); // Value being mod'd | |
| 852 Node *divisor = in(2); // Also is mask | |
| 853 | |
| 854 hook->init_req(0, x); // Add a use to x to prevent him from dying | |
| 855 // Generate code to reduce X rapidly to nearly 2^k-1. | |
| 856 for( int i = 0; i < trip_count; i++ ) { | |
| 145 | 857 Node *xl = phase->transform( new (phase->C, 3) AndINode(x,divisor) ); |
| 858 Node *xh = phase->transform( new (phase->C, 3) RShiftINode(x,phase->intcon(k)) ); // Must be signed | |
| 859 x = phase->transform( new (phase->C, 3) AddINode(xh,xl) ); | |
| 860 hook->set_req(0, x); | |
| 0 | 861 } |
| 862 | |
| 863 // Generate sign-fixup code. Was original value positive? | |
| 864 // int hack_res = (i >= 0) ? divisor : 1; | |
| 865 Node *cmp1 = phase->transform( new (phase->C, 3) CmpINode( in(1), phase->intcon(0) ) ); | |
| 866 Node *bol1 = phase->transform( new (phase->C, 2) BoolNode( cmp1, BoolTest::ge ) ); | |
| 867 Node *cmov1= phase->transform( new (phase->C, 4) CMoveINode(bol1, phase->intcon(1), divisor, TypeInt::POS) ); | |
| 868 // if( x >= hack_res ) x -= divisor; | |
| 869 Node *sub = phase->transform( new (phase->C, 3) SubINode( x, divisor ) ); | |
| 870 Node *cmp2 = phase->transform( new (phase->C, 3) CmpINode( x, cmov1 ) ); | |
| 871 Node *bol2 = phase->transform( new (phase->C, 2) BoolNode( cmp2, BoolTest::ge ) ); | |
| 872 // Convention is to not transform the return value of an Ideal | |
| 873 // since Ideal is expected to return a modified 'this' or a new node. | |
| 874 Node *cmov2= new (phase->C, 4) CMoveINode(bol2, x, sub, TypeInt::INT); | |
| 875 // cmov2 is now the mod | |
| 876 | |
| 877 // Now remove the bogus extra edges used to keep things alive | |
| 878 if (can_reshape) { | |
| 879 phase->is_IterGVN()->remove_dead_node(hook); | |
| 880 } else { | |
| 881 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
| 882 } | |
| 883 return cmov2; | |
| 884 } | |
| 885 } | |
| 886 | |
| 887 // Fell thru, the unroll case is not appropriate. Transform the modulo | |
| 888 // into a long multiply/int multiply/subtract case | |
| 889 | |
| 890 // Cannot handle mod 0, and min_jint isn't handled by the transform | |
| 891 if( con == 0 || con == min_jint ) return NULL; | |
| 892 | |
| 893 // Get the absolute value of the constant; at this point, we can use this | |
| 894 jint pos_con = (con >= 0) ? con : -con; | |
| 895 | |
| 896 // integer Mod 1 is always 0 | |
| 897 if( pos_con == 1 ) return new (phase->C, 1) ConINode(TypeInt::ZERO); | |
| 898 | |
| 899 int log2_con = -1; | |
| 900 | |
| 901 // If this is a power of two, they maybe we can mask it | |
| 902 if( is_power_of_2(pos_con) ) { | |
| 903 log2_con = log2_intptr((intptr_t)pos_con); | |
| 904 | |
| 905 const Type *dt = phase->type(in(1)); | |
| 906 const TypeInt *dti = dt->isa_int(); | |
| 907 | |
| 908 // See if this can be masked, if the dividend is non-negative | |
| 909 if( dti && dti->_lo >= 0 ) | |
| 910 return ( new (phase->C, 3) AndINode( in(1), phase->intcon( pos_con-1 ) ) ); | |
| 911 } | |
| 912 | |
| 913 // Save in(1) so that it cannot be changed or deleted | |
| 914 hook->init_req(0, in(1)); | |
| 915 | |
| 916 // Divide using the transform from DivI to MulL | |
| 145 | 917 Node *result = transform_int_divide( phase, in(1), pos_con ); |
| 918 if (result != NULL) { | |
| 919 Node *divide = phase->transform(result); | |
| 0 | 920 |
| 145 | 921 // Re-multiply, using a shift if this is a power of two |
| 922 Node *mult = NULL; | |
| 0 | 923 |
| 145 | 924 if( log2_con >= 0 ) |
| 925 mult = phase->transform( new (phase->C, 3) LShiftINode( divide, phase->intcon( log2_con ) ) ); | |
| 926 else | |
| 927 mult = phase->transform( new (phase->C, 3) MulINode( divide, phase->intcon( pos_con ) ) ); | |
| 0 | 928 |
| 145 | 929 // Finally, subtract the multiplied divided value from the original |
| 930 result = new (phase->C, 3) SubINode( in(1), mult ); | |
| 931 } | |
| 0 | 932 |
| 933 // Now remove the bogus extra edges used to keep things alive | |
| 934 if (can_reshape) { | |
| 935 phase->is_IterGVN()->remove_dead_node(hook); | |
| 936 } else { | |
| 937 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
| 938 } | |
| 939 | |
| 940 // return the value | |
| 941 return result; | |
| 942 } | |
| 943 | |
| 944 //------------------------------Value------------------------------------------ | |
| 945 const Type *ModINode::Value( PhaseTransform *phase ) const { | |
| 946 // Either input is TOP ==> the result is TOP | |
| 947 const Type *t1 = phase->type( in(1) ); | |
| 948 const Type *t2 = phase->type( in(2) ); | |
| 949 if( t1 == Type::TOP ) return Type::TOP; | |
| 950 if( t2 == Type::TOP ) return Type::TOP; | |
| 951 | |
| 952 // We always generate the dynamic check for 0. | |
| 953 // 0 MOD X is 0 | |
| 954 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; | |
| 955 // X MOD X is 0 | |
| 956 if( phase->eqv( in(1), in(2) ) ) return TypeInt::ZERO; | |
| 957 | |
| 958 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 959 const Type *bot = bottom_type(); | |
| 960 if( (t1 == bot) || (t2 == bot) || | |
| 961 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 962 return bot; | |
| 963 | |
| 964 const TypeInt *i1 = t1->is_int(); | |
| 965 const TypeInt *i2 = t2->is_int(); | |
| 966 if( !i1->is_con() || !i2->is_con() ) { | |
| 967 if( i1->_lo >= 0 && i2->_lo >= 0 ) | |
| 968 return TypeInt::POS; | |
| 969 // If both numbers are not constants, we know little. | |
| 970 return TypeInt::INT; | |
| 971 } | |
| 972 // Mod by zero? Throw exception at runtime! | |
| 973 if( !i2->get_con() ) return TypeInt::POS; | |
| 974 | |
| 975 // We must be modulo'ing 2 float constants. | |
| 976 // Check for min_jint % '-1', result is defined to be '0'. | |
| 977 if( i1->get_con() == min_jint && i2->get_con() == -1 ) | |
| 978 return TypeInt::ZERO; | |
| 979 | |
| 980 return TypeInt::make( i1->get_con() % i2->get_con() ); | |
| 981 } | |
| 982 | |
| 983 | |
| 984 //============================================================================= | |
| 985 //------------------------------Idealize--------------------------------------- | |
| 986 Node *ModLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
| 987 // Check for dead control input | |
| 305 | 988 if( in(0) && remove_dead_region(phase, can_reshape) ) return this; |
| 989 // Don't bother trying to transform a dead node | |
| 990 if( in(0) && in(0)->is_top() ) return NULL; | |
| 0 | 991 |
| 992 // Get the modulus | |
| 993 const Type *t = phase->type( in(2) ); | |
| 994 if( t == Type::TOP ) return NULL; | |
| 145 | 995 const TypeLong *tl = t->is_long(); |
| 0 | 996 |
| 997 // Check for useless control input | |
| 998 // Check for excluding mod-zero case | |
| 145 | 999 if( in(0) && (tl->_hi < 0 || tl->_lo > 0) ) { |
| 0 | 1000 set_req(0, NULL); // Yank control input |
| 1001 return this; | |
| 1002 } | |
| 1003 | |
| 1004 // See if we are MOD'ing by 2^k or 2^k-1. | |
| 145 | 1005 if( !tl->is_con() ) return NULL; |
| 1006 jlong con = tl->get_con(); | |
| 1007 | |
| 1008 Node *hook = new (phase->C, 1) Node(1); | |
| 0 | 1009 |
| 1010 // Expand mod | |
| 145 | 1011 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
|
1012 uint k = exact_log2_long(con+1); // Extract k |
| 145 | 1013 |
| 0 | 1014 // Basic algorithm by David Detlefs. See fastmod_long.java for gory details. |
| 1015 // Used to help a popular random number generator which does a long-mod | |
| 1016 // of 2^31-1 and shows up in SpecJBB and SciMark. | |
| 1017 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*/}; | |
| 1018 int trip_count = 1; | |
| 1019 if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k]; | |
| 1020 | |
| 145 | 1021 // If the unroll factor is not too large, and if conditional moves are |
| 1022 // ok, then use this case | |
| 1023 if( trip_count <= 5 && ConditionalMoveLimit != 0 ) { | |
| 1024 Node *x = in(1); // Value being mod'd | |
| 1025 Node *divisor = in(2); // Also is mask | |
| 0 | 1026 |
| 145 | 1027 hook->init_req(0, x); // Add a use to x to prevent him from dying |
| 1028 // Generate code to reduce X rapidly to nearly 2^k-1. | |
| 1029 for( int i = 0; i < trip_count; i++ ) { | |
| 0 | 1030 Node *xl = phase->transform( new (phase->C, 3) AndLNode(x,divisor) ); |
| 1031 Node *xh = phase->transform( new (phase->C, 3) RShiftLNode(x,phase->intcon(k)) ); // Must be signed | |
| 1032 x = phase->transform( new (phase->C, 3) AddLNode(xh,xl) ); | |
| 1033 hook->set_req(0, x); // Add a use to x to prevent him from dying | |
| 145 | 1034 } |
| 1035 | |
| 1036 // Generate sign-fixup code. Was original value positive? | |
| 1037 // long hack_res = (i >= 0) ? divisor : CONST64(1); | |
| 1038 Node *cmp1 = phase->transform( new (phase->C, 3) CmpLNode( in(1), phase->longcon(0) ) ); | |
| 1039 Node *bol1 = phase->transform( new (phase->C, 2) BoolNode( cmp1, BoolTest::ge ) ); | |
| 1040 Node *cmov1= phase->transform( new (phase->C, 4) CMoveLNode(bol1, phase->longcon(1), divisor, TypeLong::LONG) ); | |
| 1041 // if( x >= hack_res ) x -= divisor; | |
| 1042 Node *sub = phase->transform( new (phase->C, 3) SubLNode( x, divisor ) ); | |
| 1043 Node *cmp2 = phase->transform( new (phase->C, 3) CmpLNode( x, cmov1 ) ); | |
| 1044 Node *bol2 = phase->transform( new (phase->C, 2) BoolNode( cmp2, BoolTest::ge ) ); | |
| 1045 // Convention is to not transform the return value of an Ideal | |
| 1046 // since Ideal is expected to return a modified 'this' or a new node. | |
| 1047 Node *cmov2= new (phase->C, 4) CMoveLNode(bol2, x, sub, TypeLong::LONG); | |
| 1048 // cmov2 is now the mod | |
| 1049 | |
| 1050 // Now remove the bogus extra edges used to keep things alive | |
| 1051 if (can_reshape) { | |
| 1052 phase->is_IterGVN()->remove_dead_node(hook); | |
| 1053 } else { | |
| 1054 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
| 1055 } | |
| 1056 return cmov2; | |
| 0 | 1057 } |
| 145 | 1058 } |
| 1059 | |
| 1060 // Fell thru, the unroll case is not appropriate. Transform the modulo | |
| 1061 // into a long multiply/int multiply/subtract case | |
| 1062 | |
| 1063 // Cannot handle mod 0, and min_jint isn't handled by the transform | |
| 1064 if( con == 0 || con == min_jlong ) return NULL; | |
| 1065 | |
| 1066 // Get the absolute value of the constant; at this point, we can use this | |
| 1067 jlong pos_con = (con >= 0) ? con : -con; | |
| 1068 | |
| 1069 // integer Mod 1 is always 0 | |
| 1070 if( pos_con == 1 ) return new (phase->C, 1) ConLNode(TypeLong::ZERO); | |
| 1071 | |
| 1072 int log2_con = -1; | |
| 1073 | |
| 605 | 1074 // If this is a power of two, then maybe we can mask it |
| 145 | 1075 if( is_power_of_2_long(pos_con) ) { |
| 1076 log2_con = log2_long(pos_con); | |
| 1077 | |
| 1078 const Type *dt = phase->type(in(1)); | |
| 1079 const TypeLong *dtl = dt->isa_long(); | |
| 1080 | |
| 1081 // See if this can be masked, if the dividend is non-negative | |
| 1082 if( dtl && dtl->_lo >= 0 ) | |
| 1083 return ( new (phase->C, 3) AndLNode( in(1), phase->longcon( pos_con-1 ) ) ); | |
| 1084 } | |
| 0 | 1085 |
| 145 | 1086 // Save in(1) so that it cannot be changed or deleted |
| 1087 hook->init_req(0, in(1)); | |
| 1088 | |
| 1089 // Divide using the transform from DivI to MulL | |
| 1090 Node *result = transform_long_divide( phase, in(1), pos_con ); | |
| 1091 if (result != NULL) { | |
| 1092 Node *divide = phase->transform(result); | |
| 1093 | |
| 1094 // Re-multiply, using a shift if this is a power of two | |
| 1095 Node *mult = NULL; | |
| 1096 | |
| 1097 if( log2_con >= 0 ) | |
| 1098 mult = phase->transform( new (phase->C, 3) LShiftLNode( divide, phase->intcon( log2_con ) ) ); | |
| 1099 else | |
| 1100 mult = phase->transform( new (phase->C, 3) MulLNode( divide, phase->longcon( pos_con ) ) ); | |
| 1101 | |
| 1102 // Finally, subtract the multiplied divided value from the original | |
| 1103 result = new (phase->C, 3) SubLNode( in(1), mult ); | |
| 0 | 1104 } |
| 145 | 1105 |
| 1106 // Now remove the bogus extra edges used to keep things alive | |
| 1107 if (can_reshape) { | |
| 1108 phase->is_IterGVN()->remove_dead_node(hook); | |
| 1109 } else { | |
| 1110 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
| 1111 } | |
| 1112 | |
| 1113 // return the value | |
| 1114 return result; | |
| 0 | 1115 } |
| 1116 | |
| 1117 //------------------------------Value------------------------------------------ | |
| 1118 const Type *ModLNode::Value( PhaseTransform *phase ) const { | |
| 1119 // Either input is TOP ==> the result is TOP | |
| 1120 const Type *t1 = phase->type( in(1) ); | |
| 1121 const Type *t2 = phase->type( in(2) ); | |
| 1122 if( t1 == Type::TOP ) return Type::TOP; | |
| 1123 if( t2 == Type::TOP ) return Type::TOP; | |
| 1124 | |
| 1125 // We always generate the dynamic check for 0. | |
| 1126 // 0 MOD X is 0 | |
| 1127 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; | |
| 1128 // X MOD X is 0 | |
| 1129 if( phase->eqv( in(1), in(2) ) ) return TypeLong::ZERO; | |
| 1130 | |
| 1131 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 1132 const Type *bot = bottom_type(); | |
| 1133 if( (t1 == bot) || (t2 == bot) || | |
| 1134 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 1135 return bot; | |
| 1136 | |
| 1137 const TypeLong *i1 = t1->is_long(); | |
| 1138 const TypeLong *i2 = t2->is_long(); | |
| 1139 if( !i1->is_con() || !i2->is_con() ) { | |
| 1140 if( i1->_lo >= CONST64(0) && i2->_lo >= CONST64(0) ) | |
| 1141 return TypeLong::POS; | |
| 1142 // If both numbers are not constants, we know little. | |
| 1143 return TypeLong::LONG; | |
| 1144 } | |
| 1145 // Mod by zero? Throw exception at runtime! | |
| 1146 if( !i2->get_con() ) return TypeLong::POS; | |
| 1147 | |
| 1148 // We must be modulo'ing 2 float constants. | |
| 1149 // Check for min_jint % '-1', result is defined to be '0'. | |
| 1150 if( i1->get_con() == min_jlong && i2->get_con() == -1 ) | |
| 1151 return TypeLong::ZERO; | |
| 1152 | |
| 1153 return TypeLong::make( i1->get_con() % i2->get_con() ); | |
| 1154 } | |
| 1155 | |
| 1156 | |
| 1157 //============================================================================= | |
| 1158 //------------------------------Value------------------------------------------ | |
| 1159 const Type *ModFNode::Value( PhaseTransform *phase ) const { | |
| 1160 // Either input is TOP ==> the result is TOP | |
| 1161 const Type *t1 = phase->type( in(1) ); | |
| 1162 const Type *t2 = phase->type( in(2) ); | |
| 1163 if( t1 == Type::TOP ) return Type::TOP; | |
| 1164 if( t2 == Type::TOP ) return Type::TOP; | |
| 1165 | |
| 1166 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 1167 const Type *bot = bottom_type(); | |
| 1168 if( (t1 == bot) || (t2 == bot) || | |
| 1169 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 1170 return bot; | |
| 1171 | |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1172 // 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
|
1173 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
|
1174 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
|
1175 } |
| 0 | 1176 |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1177 float f1 = t1->getf(); |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1178 float f2 = t2->getf(); |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1179 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
|
1180 jint x2 = jint_cast(f2); |
| 0 | 1181 |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1182 // 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
|
1183 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
|
1184 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
|
1185 |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1186 // 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
|
1187 if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jint) |
| 0 | 1188 return Type::FLOAT; |
| 1189 | |
| 1190 // We must be modulo'ing 2 float constants. | |
| 1191 // 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
|
1192 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
|
1193 if ((x1 ^ xr) < 0) { |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1194 xr ^= min_jint; |
| 0 | 1195 } |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1196 |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1197 return TypeF::make(jfloat_cast(xr)); |
| 0 | 1198 } |
| 1199 | |
| 1200 | |
| 1201 //============================================================================= | |
| 1202 //------------------------------Value------------------------------------------ | |
| 1203 const Type *ModDNode::Value( PhaseTransform *phase ) const { | |
| 1204 // Either input is TOP ==> the result is TOP | |
| 1205 const Type *t1 = phase->type( in(1) ); | |
| 1206 const Type *t2 = phase->type( in(2) ); | |
| 1207 if( t1 == Type::TOP ) return Type::TOP; | |
| 1208 if( t2 == Type::TOP ) return Type::TOP; | |
| 1209 | |
| 1210 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 1211 const Type *bot = bottom_type(); | |
| 1212 if( (t1 == bot) || (t2 == bot) || | |
| 1213 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 1214 return bot; | |
| 1215 | |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1216 // 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
|
1217 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
|
1218 return Type::DOUBLE; // note: x%x can be either NaN or 0 |
| 0 | 1219 } |
| 1220 | |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1221 double f1 = t1->getd(); |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1222 double f2 = t2->getd(); |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1223 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
|
1224 jlong x2 = jlong_cast(f2); |
| 0 | 1225 |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1226 // 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
|
1227 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
|
1228 if (g_isnan(f2)) return t2; |
| 0 | 1229 |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1230 // 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
|
1231 if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jlong) |
| 0 | 1232 return Type::DOUBLE; |
| 1233 | |
| 1234 // 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
|
1235 // 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
|
1236 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
|
1237 if ((x1 ^ xr) < 0) { |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1238 xr ^= min_jlong; |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1239 } |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1240 |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1241 return TypeD::make(jdouble_cast(xr)); |
| 0 | 1242 } |
| 1243 | |
| 1244 //============================================================================= | |
| 1245 | |
| 1246 DivModNode::DivModNode( Node *c, Node *dividend, Node *divisor ) : MultiNode(3) { | |
| 1247 init_req(0, c); | |
| 1248 init_req(1, dividend); | |
| 1249 init_req(2, divisor); | |
| 1250 } | |
| 1251 | |
| 1252 //------------------------------make------------------------------------------ | |
| 1253 DivModINode* DivModINode::make(Compile* C, Node* div_or_mod) { | |
| 1254 Node* n = div_or_mod; | |
| 1255 assert(n->Opcode() == Op_DivI || n->Opcode() == Op_ModI, | |
| 1256 "only div or mod input pattern accepted"); | |
| 1257 | |
| 1258 DivModINode* divmod = new (C, 3) DivModINode(n->in(0), n->in(1), n->in(2)); | |
| 1259 Node* dproj = new (C, 1) ProjNode(divmod, DivModNode::div_proj_num); | |
| 1260 Node* mproj = new (C, 1) ProjNode(divmod, DivModNode::mod_proj_num); | |
| 1261 return divmod; | |
| 1262 } | |
| 1263 | |
| 1264 //------------------------------make------------------------------------------ | |
| 1265 DivModLNode* DivModLNode::make(Compile* C, Node* div_or_mod) { | |
| 1266 Node* n = div_or_mod; | |
| 1267 assert(n->Opcode() == Op_DivL || n->Opcode() == Op_ModL, | |
| 1268 "only div or mod input pattern accepted"); | |
| 1269 | |
| 1270 DivModLNode* divmod = new (C, 3) DivModLNode(n->in(0), n->in(1), n->in(2)); | |
| 1271 Node* dproj = new (C, 1) ProjNode(divmod, DivModNode::div_proj_num); | |
| 1272 Node* mproj = new (C, 1) ProjNode(divmod, DivModNode::mod_proj_num); | |
| 1273 return divmod; | |
| 1274 } | |
| 1275 | |
| 1276 //------------------------------match------------------------------------------ | |
| 1277 // return result(s) along with their RegMask info | |
| 1278 Node *DivModINode::match( const ProjNode *proj, const Matcher *match ) { | |
| 1279 uint ideal_reg = proj->ideal_reg(); | |
| 1280 RegMask rm; | |
| 1281 if (proj->_con == div_proj_num) { | |
| 1282 rm = match->divI_proj_mask(); | |
| 1283 } else { | |
| 1284 assert(proj->_con == mod_proj_num, "must be div or mod projection"); | |
| 1285 rm = match->modI_proj_mask(); | |
| 1286 } | |
| 1287 return new (match->C, 1)MachProjNode(this, proj->_con, rm, ideal_reg); | |
| 1288 } | |
| 1289 | |
| 1290 | |
| 1291 //------------------------------match------------------------------------------ | |
| 1292 // return result(s) along with their RegMask info | |
| 1293 Node *DivModLNode::match( const ProjNode *proj, const Matcher *match ) { | |
| 1294 uint ideal_reg = proj->ideal_reg(); | |
| 1295 RegMask rm; | |
| 1296 if (proj->_con == div_proj_num) { | |
| 1297 rm = match->divL_proj_mask(); | |
| 1298 } else { | |
| 1299 assert(proj->_con == mod_proj_num, "must be div or mod projection"); | |
| 1300 rm = match->modL_proj_mask(); | |
| 1301 } | |
| 1302 return new (match->C, 1)MachProjNode(this, proj->_con, rm, ideal_reg); | |
| 1303 } |