Mercurial > hg > graal-compiler
annotate src/share/vm/opto/divnode.cpp @ 1914:ae065c367d93
6987135: Performance regression on Intel platform with 32-bits edition between 6u13 and 6u14.
Summary: Use hardware DIV instruction for long division by constant when it is faster than code with multiply.
Reviewed-by: never
| author | kvn |
|---|---|
| date | Tue, 02 Nov 2010 09:00:37 -0700 |
| parents | c18cbe5936b8 |
| children | f95d63e2154a |
| rev | line source |
|---|---|
| 0 | 1 /* |
|
1914
ae065c367d93
6987135: Performance regression on Intel platform with 32-bits edition between 6u13 and 6u14.
kvn
parents:
1552
diff
changeset
|
2 * Copyright (c) 1997, 2010, 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 } | |
|
1914
ae065c367d93
6987135: Performance regression on Intel platform with 32-bits edition between 6u13 and 6u14.
kvn
parents:
1552
diff
changeset
|
391 } else if ( !Matcher::use_asm_for_ldiv_by_con(d) ) { // Use hardware DIV instruction when |
|
ae065c367d93
6987135: Performance regression on Intel platform with 32-bits edition between 6u13 and 6u14.
kvn
parents:
1552
diff
changeset
|
392 // it is faster than code generated below. |
| 145 | 393 // Attempt the jlong constant divide -> multiply transform found in |
| 394 // "Division by Invariant Integers using Multiplication" | |
| 395 // by Granlund and Montgomery | |
| 396 // See also "Hacker's Delight", chapter 10 by Warren. | |
| 397 | |
| 398 jlong magic_const; | |
| 399 jint shift_const; | |
| 400 if (magic_long_divide_constants(d, magic_const, shift_const)) { | |
| 401 // Compute the high half of the dividend x magic multiplication | |
| 402 Node *mul_hi = phase->transform(long_by_long_mulhi(phase, dividend, magic_const)); | |
| 403 | |
| 404 // The high half of the 128-bit multiply is computed. | |
| 405 if (magic_const < 0) { | |
| 406 // The magic multiplier is too large for a 64 bit constant. We've adjusted | |
| 407 // it down by 2^64, but have to add 1 dividend back in after the multiplication. | |
| 408 // This handles the "overflow" case described by Granlund and Montgomery. | |
| 409 mul_hi = phase->transform(new (phase->C, 3) AddLNode(dividend, mul_hi)); | |
| 410 } | |
| 411 | |
| 412 // Shift over the (adjusted) mulhi | |
| 413 if (shift_const != 0) { | |
| 414 mul_hi = phase->transform(new (phase->C, 3) RShiftLNode(mul_hi, phase->intcon(shift_const))); | |
| 415 } | |
| 416 | |
| 417 // Get a 0 or -1 from the sign of the dividend. | |
| 418 Node *addend0 = mul_hi; | |
| 419 Node *addend1 = phase->transform(new (phase->C, 3) RShiftLNode(dividend, phase->intcon(N-1))); | |
| 420 | |
| 421 // If the divisor is negative, swap the order of the input addends; | |
| 422 // this has the effect of negating the quotient. | |
| 423 if (!d_pos) { | |
| 424 Node *temp = addend0; addend0 = addend1; addend1 = temp; | |
| 425 } | |
| 426 | |
| 427 // Adjust the final quotient by subtracting -1 (adding 1) | |
| 428 // from the mul_hi. | |
| 429 q = new (phase->C, 3) SubLNode(addend0, addend1); | |
| 430 } | |
| 0 | 431 } |
| 432 | |
| 145 | 433 return q; |
| 0 | 434 } |
| 435 | |
| 436 //============================================================================= | |
| 437 //------------------------------Identity--------------------------------------- | |
| 438 // If the divisor is 1, we are an identity on the dividend. | |
| 439 Node *DivINode::Identity( PhaseTransform *phase ) { | |
| 440 return (phase->type( in(2) )->higher_equal(TypeInt::ONE)) ? in(1) : this; | |
| 441 } | |
| 442 | |
| 443 //------------------------------Idealize--------------------------------------- | |
| 444 // Divides can be changed to multiplies and/or shifts | |
| 445 Node *DivINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
| 446 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
| 305 | 447 // Don't bother trying to transform a dead node |
| 448 if( in(0) && in(0)->is_top() ) return NULL; | |
| 0 | 449 |
| 450 const Type *t = phase->type( in(2) ); | |
| 451 if( t == TypeInt::ONE ) // Identity? | |
| 452 return NULL; // Skip it | |
| 453 | |
| 454 const TypeInt *ti = t->isa_int(); | |
| 455 if( !ti ) return NULL; | |
| 456 if( !ti->is_con() ) return NULL; | |
| 145 | 457 jint i = ti->get_con(); // Get divisor |
| 0 | 458 |
| 459 if (i == 0) return NULL; // Dividing by zero constant does not idealize | |
| 460 | |
| 461 set_req(0,NULL); // Dividing by a not-zero constant; no faulting | |
| 462 | |
| 463 // Dividing by MININT does not optimize as a power-of-2 shift. | |
| 464 if( i == min_jint ) return NULL; | |
| 465 | |
| 145 | 466 return transform_int_divide( phase, in(1), i ); |
| 0 | 467 } |
| 468 | |
| 469 //------------------------------Value------------------------------------------ | |
| 470 // A DivINode divides its inputs. The third input is a Control input, used to | |
| 471 // prevent hoisting the divide above an unsafe test. | |
| 472 const Type *DivINode::Value( PhaseTransform *phase ) const { | |
| 473 // Either input is TOP ==> the result is TOP | |
| 474 const Type *t1 = phase->type( in(1) ); | |
| 475 const Type *t2 = phase->type( in(2) ); | |
| 476 if( t1 == Type::TOP ) return Type::TOP; | |
| 477 if( t2 == Type::TOP ) return Type::TOP; | |
| 478 | |
| 479 // x/x == 1 since we always generate the dynamic divisor check for 0. | |
| 480 if( phase->eqv( in(1), in(2) ) ) | |
| 481 return TypeInt::ONE; | |
| 482 | |
| 483 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 484 const Type *bot = bottom_type(); | |
| 485 if( (t1 == bot) || (t2 == bot) || | |
| 486 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 487 return bot; | |
| 488 | |
| 489 // Divide the two numbers. We approximate. | |
| 490 // If divisor is a constant and not zero | |
| 491 const TypeInt *i1 = t1->is_int(); | |
| 492 const TypeInt *i2 = t2->is_int(); | |
| 493 int widen = MAX2(i1->_widen, i2->_widen); | |
| 494 | |
| 495 if( i2->is_con() && i2->get_con() != 0 ) { | |
| 496 int32 d = i2->get_con(); // Divisor | |
| 497 jint lo, hi; | |
| 498 if( d >= 0 ) { | |
| 499 lo = i1->_lo/d; | |
| 500 hi = i1->_hi/d; | |
| 501 } else { | |
| 502 if( d == -1 && i1->_lo == min_jint ) { | |
| 503 // 'min_jint/-1' throws arithmetic exception during compilation | |
| 504 lo = min_jint; | |
| 505 // do not support holes, 'hi' must go to either min_jint or max_jint: | |
| 506 // [min_jint, -10]/[-1,-1] ==> [min_jint] UNION [10,max_jint] | |
| 507 hi = i1->_hi == min_jint ? min_jint : max_jint; | |
| 508 } else { | |
| 509 lo = i1->_hi/d; | |
| 510 hi = i1->_lo/d; | |
| 511 } | |
| 512 } | |
| 513 return TypeInt::make(lo, hi, widen); | |
| 514 } | |
| 515 | |
| 516 // If the dividend is a constant | |
| 517 if( i1->is_con() ) { | |
| 518 int32 d = i1->get_con(); | |
| 519 if( d < 0 ) { | |
| 520 if( d == min_jint ) { | |
| 521 // (-min_jint) == min_jint == (min_jint / -1) | |
| 522 return TypeInt::make(min_jint, max_jint/2 + 1, widen); | |
| 523 } else { | |
| 524 return TypeInt::make(d, -d, widen); | |
| 525 } | |
| 526 } | |
| 527 return TypeInt::make(-d, d, widen); | |
| 528 } | |
| 529 | |
| 530 // Otherwise we give up all hope | |
| 531 return TypeInt::INT; | |
| 532 } | |
| 533 | |
| 534 | |
| 535 //============================================================================= | |
| 536 //------------------------------Identity--------------------------------------- | |
| 537 // If the divisor is 1, we are an identity on the dividend. | |
| 538 Node *DivLNode::Identity( PhaseTransform *phase ) { | |
| 539 return (phase->type( in(2) )->higher_equal(TypeLong::ONE)) ? in(1) : this; | |
| 540 } | |
| 541 | |
| 542 //------------------------------Idealize--------------------------------------- | |
| 543 // Dividing by a power of 2 is a shift. | |
| 544 Node *DivLNode::Ideal( PhaseGVN *phase, bool can_reshape) { | |
| 545 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
| 305 | 546 // Don't bother trying to transform a dead node |
| 547 if( in(0) && in(0)->is_top() ) return NULL; | |
| 0 | 548 |
| 549 const Type *t = phase->type( in(2) ); | |
| 145 | 550 if( t == TypeLong::ONE ) // Identity? |
| 0 | 551 return NULL; // Skip it |
| 552 | |
| 145 | 553 const TypeLong *tl = t->isa_long(); |
| 554 if( !tl ) return NULL; | |
| 555 if( !tl->is_con() ) return NULL; | |
| 556 jlong l = tl->get_con(); // Get divisor | |
| 557 | |
| 558 if (l == 0) return NULL; // Dividing by zero constant does not idealize | |
| 559 | |
| 560 set_req(0,NULL); // Dividing by a not-zero constant; no faulting | |
| 0 | 561 |
|
1914
ae065c367d93
6987135: Performance regression on Intel platform with 32-bits edition between 6u13 and 6u14.
kvn
parents:
1552
diff
changeset
|
562 // Dividing by MINLONG does not optimize as a power-of-2 shift. |
| 145 | 563 if( l == min_jlong ) return NULL; |
| 0 | 564 |
| 145 | 565 return transform_long_divide( phase, in(1), l ); |
| 0 | 566 } |
| 567 | |
| 568 //------------------------------Value------------------------------------------ | |
| 569 // A DivLNode divides its inputs. The third input is a Control input, used to | |
| 570 // prevent hoisting the divide above an unsafe test. | |
| 571 const Type *DivLNode::Value( PhaseTransform *phase ) const { | |
| 572 // Either input is TOP ==> the result is TOP | |
| 573 const Type *t1 = phase->type( in(1) ); | |
| 574 const Type *t2 = phase->type( in(2) ); | |
| 575 if( t1 == Type::TOP ) return Type::TOP; | |
| 576 if( t2 == Type::TOP ) return Type::TOP; | |
| 577 | |
| 578 // x/x == 1 since we always generate the dynamic divisor check for 0. | |
| 579 if( phase->eqv( in(1), in(2) ) ) | |
| 580 return TypeLong::ONE; | |
| 581 | |
| 582 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 583 const Type *bot = bottom_type(); | |
| 584 if( (t1 == bot) || (t2 == bot) || | |
| 585 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 586 return bot; | |
| 587 | |
| 588 // Divide the two numbers. We approximate. | |
| 589 // If divisor is a constant and not zero | |
| 590 const TypeLong *i1 = t1->is_long(); | |
| 591 const TypeLong *i2 = t2->is_long(); | |
| 592 int widen = MAX2(i1->_widen, i2->_widen); | |
| 593 | |
| 594 if( i2->is_con() && i2->get_con() != 0 ) { | |
| 595 jlong d = i2->get_con(); // Divisor | |
| 596 jlong lo, hi; | |
| 597 if( d >= 0 ) { | |
| 598 lo = i1->_lo/d; | |
| 599 hi = i1->_hi/d; | |
| 600 } else { | |
| 601 if( d == CONST64(-1) && i1->_lo == min_jlong ) { | |
| 602 // 'min_jlong/-1' throws arithmetic exception during compilation | |
| 603 lo = min_jlong; | |
| 604 // do not support holes, 'hi' must go to either min_jlong or max_jlong: | |
| 605 // [min_jlong, -10]/[-1,-1] ==> [min_jlong] UNION [10,max_jlong] | |
| 606 hi = i1->_hi == min_jlong ? min_jlong : max_jlong; | |
| 607 } else { | |
| 608 lo = i1->_hi/d; | |
| 609 hi = i1->_lo/d; | |
| 610 } | |
| 611 } | |
| 612 return TypeLong::make(lo, hi, widen); | |
| 613 } | |
| 614 | |
| 615 // If the dividend is a constant | |
| 616 if( i1->is_con() ) { | |
| 617 jlong d = i1->get_con(); | |
| 618 if( d < 0 ) { | |
| 619 if( d == min_jlong ) { | |
| 620 // (-min_jlong) == min_jlong == (min_jlong / -1) | |
| 621 return TypeLong::make(min_jlong, max_jlong/2 + 1, widen); | |
| 622 } else { | |
| 623 return TypeLong::make(d, -d, widen); | |
| 624 } | |
| 625 } | |
| 626 return TypeLong::make(-d, d, widen); | |
| 627 } | |
| 628 | |
| 629 // Otherwise we give up all hope | |
| 630 return TypeLong::LONG; | |
| 631 } | |
| 632 | |
| 633 | |
| 634 //============================================================================= | |
| 635 //------------------------------Value------------------------------------------ | |
| 636 // An DivFNode divides its inputs. The third input is a Control input, used to | |
| 637 // prevent hoisting the divide above an unsafe test. | |
| 638 const Type *DivFNode::Value( PhaseTransform *phase ) const { | |
| 639 // Either input is TOP ==> the result is TOP | |
| 640 const Type *t1 = phase->type( in(1) ); | |
| 641 const Type *t2 = phase->type( in(2) ); | |
| 642 if( t1 == Type::TOP ) return Type::TOP; | |
| 643 if( t2 == Type::TOP ) return Type::TOP; | |
| 644 | |
| 645 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 646 const Type *bot = bottom_type(); | |
| 647 if( (t1 == bot) || (t2 == bot) || | |
| 648 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 649 return bot; | |
| 650 | |
| 651 // x/x == 1, we ignore 0/0. | |
| 652 // 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
|
653 // Does not work for variables because of NaN's |
| 0 | 654 if( phase->eqv( in(1), in(2) ) && t1->base() == Type::FloatCon) |
| 655 if (!g_isnan(t1->getf()) && g_isfinite(t1->getf()) && t1->getf() != 0.0) // could be negative ZERO or NaN | |
| 656 return TypeF::ONE; | |
| 657 | |
| 658 if( t2 == TypeF::ONE ) | |
| 659 return t1; | |
| 660 | |
| 661 // If divisor is a constant and not zero, divide them numbers | |
| 662 if( t1->base() == Type::FloatCon && | |
| 663 t2->base() == Type::FloatCon && | |
| 664 t2->getf() != 0.0 ) // could be negative zero | |
| 665 return TypeF::make( t1->getf()/t2->getf() ); | |
| 666 | |
| 667 // If the dividend is a constant zero | |
| 668 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
| 669 // Test TypeF::ZERO is not sufficient as it could be negative zero | |
| 670 | |
| 671 if( t1 == TypeF::ZERO && !g_isnan(t2->getf()) && t2->getf() != 0.0 ) | |
| 672 return TypeF::ZERO; | |
| 673 | |
| 674 // Otherwise we give up all hope | |
| 675 return Type::FLOAT; | |
| 676 } | |
| 677 | |
| 678 //------------------------------isA_Copy--------------------------------------- | |
| 679 // Dividing by self is 1. | |
| 680 // If the divisor is 1, we are an identity on the dividend. | |
| 681 Node *DivFNode::Identity( PhaseTransform *phase ) { | |
| 682 return (phase->type( in(2) ) == TypeF::ONE) ? in(1) : this; | |
| 683 } | |
| 684 | |
| 685 | |
| 686 //------------------------------Idealize--------------------------------------- | |
| 687 Node *DivFNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
| 688 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
| 305 | 689 // Don't bother trying to transform a dead node |
| 690 if( in(0) && in(0)->is_top() ) return NULL; | |
| 0 | 691 |
| 692 const Type *t2 = phase->type( in(2) ); | |
| 693 if( t2 == TypeF::ONE ) // Identity? | |
| 694 return NULL; // Skip it | |
| 695 | |
| 696 const TypeF *tf = t2->isa_float_constant(); | |
| 697 if( !tf ) return NULL; | |
| 698 if( tf->base() != Type::FloatCon ) return NULL; | |
| 699 | |
| 700 // Check for out of range values | |
| 701 if( tf->is_nan() || !tf->is_finite() ) return NULL; | |
| 702 | |
| 703 // Get the value | |
| 704 float f = tf->getf(); | |
| 705 int exp; | |
| 706 | |
| 707 // Only for special case of dividing by a power of 2 | |
| 708 if( frexp((double)f, &exp) != 0.5 ) return NULL; | |
| 709 | |
| 710 // Limit the range of acceptable exponents | |
| 711 if( exp < -126 || exp > 126 ) return NULL; | |
| 712 | |
| 713 // Compute the reciprocal | |
| 714 float reciprocal = ((float)1.0) / f; | |
| 715 | |
| 716 assert( frexp((double)reciprocal, &exp) == 0.5, "reciprocal should be power of 2" ); | |
| 717 | |
| 718 // return multiplication by the reciprocal | |
| 719 return (new (phase->C, 3) MulFNode(in(1), phase->makecon(TypeF::make(reciprocal)))); | |
| 720 } | |
| 721 | |
| 722 //============================================================================= | |
| 723 //------------------------------Value------------------------------------------ | |
| 724 // 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
|
725 // prevent hoisting the divide above an unsafe test. |
| 0 | 726 const Type *DivDNode::Value( PhaseTransform *phase ) const { |
| 727 // Either input is TOP ==> the result is TOP | |
| 728 const Type *t1 = phase->type( in(1) ); | |
| 729 const Type *t2 = phase->type( in(2) ); | |
| 730 if( t1 == Type::TOP ) return Type::TOP; | |
| 731 if( t2 == Type::TOP ) return Type::TOP; | |
| 732 | |
| 733 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 734 const Type *bot = bottom_type(); | |
| 735 if( (t1 == bot) || (t2 == bot) || | |
| 736 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 737 return bot; | |
| 738 | |
| 739 // x/x == 1, we ignore 0/0. | |
| 740 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
| 741 // Does not work for variables because of NaN's | |
| 742 if( phase->eqv( in(1), in(2) ) && t1->base() == Type::DoubleCon) | |
| 743 if (!g_isnan(t1->getd()) && g_isfinite(t1->getd()) && t1->getd() != 0.0) // could be negative ZERO or NaN | |
| 744 return TypeD::ONE; | |
| 745 | |
| 746 if( t2 == TypeD::ONE ) | |
| 747 return t1; | |
| 748 | |
|
404
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
749 #if defined(IA32) |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
750 if (!phase->C->method()->is_strict()) |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
751 // 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
|
752 // division with round-to-zero on this platform. |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
753 #endif |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
754 { |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
755 // 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
|
756 if( t1->base() == Type::DoubleCon && |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
757 t2->base() == Type::DoubleCon && |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
758 t2->getd() != 0.0 ) // could be negative zero |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
759 return TypeD::make( t1->getd()/t2->getd() ); |
|
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
400
diff
changeset
|
760 } |
| 0 | 761 |
| 762 // If the dividend is a constant zero | |
| 763 // Note: if t1 and t2 are zero then result is NaN (JVMS page 213) | |
| 764 // Test TypeF::ZERO is not sufficient as it could be negative zero | |
| 765 if( t1 == TypeD::ZERO && !g_isnan(t2->getd()) && t2->getd() != 0.0 ) | |
| 766 return TypeD::ZERO; | |
| 767 | |
| 768 // Otherwise we give up all hope | |
| 769 return Type::DOUBLE; | |
| 770 } | |
| 771 | |
| 772 | |
| 773 //------------------------------isA_Copy--------------------------------------- | |
| 774 // Dividing by self is 1. | |
| 775 // If the divisor is 1, we are an identity on the dividend. | |
| 776 Node *DivDNode::Identity( PhaseTransform *phase ) { | |
| 777 return (phase->type( in(2) ) == TypeD::ONE) ? in(1) : this; | |
| 778 } | |
| 779 | |
| 780 //------------------------------Idealize--------------------------------------- | |
| 781 Node *DivDNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
| 782 if (in(0) && remove_dead_region(phase, can_reshape)) return this; | |
| 305 | 783 // Don't bother trying to transform a dead node |
| 784 if( in(0) && in(0)->is_top() ) return NULL; | |
| 0 | 785 |
| 786 const Type *t2 = phase->type( in(2) ); | |
| 787 if( t2 == TypeD::ONE ) // Identity? | |
| 788 return NULL; // Skip it | |
| 789 | |
| 790 const TypeD *td = t2->isa_double_constant(); | |
| 791 if( !td ) return NULL; | |
| 792 if( td->base() != Type::DoubleCon ) return NULL; | |
| 793 | |
| 794 // Check for out of range values | |
| 795 if( td->is_nan() || !td->is_finite() ) return NULL; | |
| 796 | |
| 797 // Get the value | |
| 798 double d = td->getd(); | |
| 799 int exp; | |
| 800 | |
| 801 // Only for special case of dividing by a power of 2 | |
| 802 if( frexp(d, &exp) != 0.5 ) return NULL; | |
| 803 | |
| 804 // Limit the range of acceptable exponents | |
| 805 if( exp < -1021 || exp > 1022 ) return NULL; | |
| 806 | |
| 807 // Compute the reciprocal | |
| 808 double reciprocal = 1.0 / d; | |
| 809 | |
| 810 assert( frexp(reciprocal, &exp) == 0.5, "reciprocal should be power of 2" ); | |
| 811 | |
| 812 // return multiplication by the reciprocal | |
| 813 return (new (phase->C, 3) MulDNode(in(1), phase->makecon(TypeD::make(reciprocal)))); | |
| 814 } | |
| 815 | |
| 816 //============================================================================= | |
| 817 //------------------------------Idealize--------------------------------------- | |
| 818 Node *ModINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
| 819 // Check for dead control input | |
| 305 | 820 if( in(0) && remove_dead_region(phase, can_reshape) ) return this; |
| 821 // Don't bother trying to transform a dead node | |
| 822 if( in(0) && in(0)->is_top() ) return NULL; | |
| 0 | 823 |
| 824 // Get the modulus | |
| 825 const Type *t = phase->type( in(2) ); | |
| 826 if( t == Type::TOP ) return NULL; | |
| 827 const TypeInt *ti = t->is_int(); | |
| 828 | |
| 829 // Check for useless control input | |
| 830 // Check for excluding mod-zero case | |
| 831 if( in(0) && (ti->_hi < 0 || ti->_lo > 0) ) { | |
| 832 set_req(0, NULL); // Yank control input | |
| 833 return this; | |
| 834 } | |
| 835 | |
| 836 // See if we are MOD'ing by 2^k or 2^k-1. | |
| 837 if( !ti->is_con() ) return NULL; | |
| 838 jint con = ti->get_con(); | |
| 839 | |
| 840 Node *hook = new (phase->C, 1) Node(1); | |
| 841 | |
| 842 // First, special check for modulo 2^k-1 | |
| 843 if( con >= 0 && con < max_jint && is_power_of_2(con+1) ) { | |
| 844 uint k = exact_log2(con+1); // Extract k | |
| 845 | |
| 846 // Basic algorithm by David Detlefs. See fastmod_int.java for gory details. | |
| 847 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*/}; | |
| 848 int trip_count = 1; | |
| 849 if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k]; | |
| 850 | |
| 851 // If the unroll factor is not too large, and if conditional moves are | |
| 852 // ok, then use this case | |
| 853 if( trip_count <= 5 && ConditionalMoveLimit != 0 ) { | |
| 854 Node *x = in(1); // Value being mod'd | |
| 855 Node *divisor = in(2); // Also is mask | |
| 856 | |
| 857 hook->init_req(0, x); // Add a use to x to prevent him from dying | |
| 858 // Generate code to reduce X rapidly to nearly 2^k-1. | |
| 859 for( int i = 0; i < trip_count; i++ ) { | |
| 145 | 860 Node *xl = phase->transform( new (phase->C, 3) AndINode(x,divisor) ); |
| 861 Node *xh = phase->transform( new (phase->C, 3) RShiftINode(x,phase->intcon(k)) ); // Must be signed | |
| 862 x = phase->transform( new (phase->C, 3) AddINode(xh,xl) ); | |
| 863 hook->set_req(0, x); | |
| 0 | 864 } |
| 865 | |
| 866 // Generate sign-fixup code. Was original value positive? | |
| 867 // int hack_res = (i >= 0) ? divisor : 1; | |
| 868 Node *cmp1 = phase->transform( new (phase->C, 3) CmpINode( in(1), phase->intcon(0) ) ); | |
| 869 Node *bol1 = phase->transform( new (phase->C, 2) BoolNode( cmp1, BoolTest::ge ) ); | |
| 870 Node *cmov1= phase->transform( new (phase->C, 4) CMoveINode(bol1, phase->intcon(1), divisor, TypeInt::POS) ); | |
| 871 // if( x >= hack_res ) x -= divisor; | |
| 872 Node *sub = phase->transform( new (phase->C, 3) SubINode( x, divisor ) ); | |
| 873 Node *cmp2 = phase->transform( new (phase->C, 3) CmpINode( x, cmov1 ) ); | |
| 874 Node *bol2 = phase->transform( new (phase->C, 2) BoolNode( cmp2, BoolTest::ge ) ); | |
| 875 // Convention is to not transform the return value of an Ideal | |
| 876 // since Ideal is expected to return a modified 'this' or a new node. | |
| 877 Node *cmov2= new (phase->C, 4) CMoveINode(bol2, x, sub, TypeInt::INT); | |
| 878 // cmov2 is now the mod | |
| 879 | |
| 880 // Now remove the bogus extra edges used to keep things alive | |
| 881 if (can_reshape) { | |
| 882 phase->is_IterGVN()->remove_dead_node(hook); | |
| 883 } else { | |
| 884 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
| 885 } | |
| 886 return cmov2; | |
| 887 } | |
| 888 } | |
| 889 | |
| 890 // Fell thru, the unroll case is not appropriate. Transform the modulo | |
| 891 // into a long multiply/int multiply/subtract case | |
| 892 | |
| 893 // Cannot handle mod 0, and min_jint isn't handled by the transform | |
| 894 if( con == 0 || con == min_jint ) return NULL; | |
| 895 | |
| 896 // Get the absolute value of the constant; at this point, we can use this | |
| 897 jint pos_con = (con >= 0) ? con : -con; | |
| 898 | |
| 899 // integer Mod 1 is always 0 | |
| 900 if( pos_con == 1 ) return new (phase->C, 1) ConINode(TypeInt::ZERO); | |
| 901 | |
| 902 int log2_con = -1; | |
| 903 | |
| 904 // If this is a power of two, they maybe we can mask it | |
| 905 if( is_power_of_2(pos_con) ) { | |
| 906 log2_con = log2_intptr((intptr_t)pos_con); | |
| 907 | |
| 908 const Type *dt = phase->type(in(1)); | |
| 909 const TypeInt *dti = dt->isa_int(); | |
| 910 | |
| 911 // See if this can be masked, if the dividend is non-negative | |
| 912 if( dti && dti->_lo >= 0 ) | |
| 913 return ( new (phase->C, 3) AndINode( in(1), phase->intcon( pos_con-1 ) ) ); | |
| 914 } | |
| 915 | |
| 916 // Save in(1) so that it cannot be changed or deleted | |
| 917 hook->init_req(0, in(1)); | |
| 918 | |
| 919 // Divide using the transform from DivI to MulL | |
| 145 | 920 Node *result = transform_int_divide( phase, in(1), pos_con ); |
| 921 if (result != NULL) { | |
| 922 Node *divide = phase->transform(result); | |
| 0 | 923 |
| 145 | 924 // Re-multiply, using a shift if this is a power of two |
| 925 Node *mult = NULL; | |
| 0 | 926 |
| 145 | 927 if( log2_con >= 0 ) |
| 928 mult = phase->transform( new (phase->C, 3) LShiftINode( divide, phase->intcon( log2_con ) ) ); | |
| 929 else | |
| 930 mult = phase->transform( new (phase->C, 3) MulINode( divide, phase->intcon( pos_con ) ) ); | |
| 0 | 931 |
| 145 | 932 // Finally, subtract the multiplied divided value from the original |
| 933 result = new (phase->C, 3) SubINode( in(1), mult ); | |
| 934 } | |
| 0 | 935 |
| 936 // Now remove the bogus extra edges used to keep things alive | |
| 937 if (can_reshape) { | |
| 938 phase->is_IterGVN()->remove_dead_node(hook); | |
| 939 } else { | |
| 940 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
| 941 } | |
| 942 | |
| 943 // return the value | |
| 944 return result; | |
| 945 } | |
| 946 | |
| 947 //------------------------------Value------------------------------------------ | |
| 948 const Type *ModINode::Value( PhaseTransform *phase ) const { | |
| 949 // Either input is TOP ==> the result is TOP | |
| 950 const Type *t1 = phase->type( in(1) ); | |
| 951 const Type *t2 = phase->type( in(2) ); | |
| 952 if( t1 == Type::TOP ) return Type::TOP; | |
| 953 if( t2 == Type::TOP ) return Type::TOP; | |
| 954 | |
| 955 // We always generate the dynamic check for 0. | |
| 956 // 0 MOD X is 0 | |
| 957 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; | |
| 958 // X MOD X is 0 | |
| 959 if( phase->eqv( in(1), in(2) ) ) return TypeInt::ZERO; | |
| 960 | |
| 961 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 962 const Type *bot = bottom_type(); | |
| 963 if( (t1 == bot) || (t2 == bot) || | |
| 964 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 965 return bot; | |
| 966 | |
| 967 const TypeInt *i1 = t1->is_int(); | |
| 968 const TypeInt *i2 = t2->is_int(); | |
| 969 if( !i1->is_con() || !i2->is_con() ) { | |
| 970 if( i1->_lo >= 0 && i2->_lo >= 0 ) | |
| 971 return TypeInt::POS; | |
| 972 // If both numbers are not constants, we know little. | |
| 973 return TypeInt::INT; | |
| 974 } | |
| 975 // Mod by zero? Throw exception at runtime! | |
| 976 if( !i2->get_con() ) return TypeInt::POS; | |
| 977 | |
| 978 // We must be modulo'ing 2 float constants. | |
| 979 // Check for min_jint % '-1', result is defined to be '0'. | |
| 980 if( i1->get_con() == min_jint && i2->get_con() == -1 ) | |
| 981 return TypeInt::ZERO; | |
| 982 | |
| 983 return TypeInt::make( i1->get_con() % i2->get_con() ); | |
| 984 } | |
| 985 | |
| 986 | |
| 987 //============================================================================= | |
| 988 //------------------------------Idealize--------------------------------------- | |
| 989 Node *ModLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
| 990 // Check for dead control input | |
| 305 | 991 if( in(0) && remove_dead_region(phase, can_reshape) ) return this; |
| 992 // Don't bother trying to transform a dead node | |
| 993 if( in(0) && in(0)->is_top() ) return NULL; | |
| 0 | 994 |
| 995 // Get the modulus | |
| 996 const Type *t = phase->type( in(2) ); | |
| 997 if( t == Type::TOP ) return NULL; | |
| 145 | 998 const TypeLong *tl = t->is_long(); |
| 0 | 999 |
| 1000 // Check for useless control input | |
| 1001 // Check for excluding mod-zero case | |
| 145 | 1002 if( in(0) && (tl->_hi < 0 || tl->_lo > 0) ) { |
| 0 | 1003 set_req(0, NULL); // Yank control input |
| 1004 return this; | |
| 1005 } | |
| 1006 | |
| 1007 // See if we are MOD'ing by 2^k or 2^k-1. | |
| 145 | 1008 if( !tl->is_con() ) return NULL; |
| 1009 jlong con = tl->get_con(); | |
| 1010 | |
| 1011 Node *hook = new (phase->C, 1) Node(1); | |
| 0 | 1012 |
| 1013 // Expand mod | |
| 145 | 1014 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
|
1015 uint k = exact_log2_long(con+1); // Extract k |
| 145 | 1016 |
| 0 | 1017 // Basic algorithm by David Detlefs. See fastmod_long.java for gory details. |
| 1018 // Used to help a popular random number generator which does a long-mod | |
| 1019 // of 2^31-1 and shows up in SpecJBB and SciMark. | |
| 1020 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*/}; | |
| 1021 int trip_count = 1; | |
| 1022 if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k]; | |
| 1023 | |
| 145 | 1024 // If the unroll factor is not too large, and if conditional moves are |
| 1025 // ok, then use this case | |
| 1026 if( trip_count <= 5 && ConditionalMoveLimit != 0 ) { | |
| 1027 Node *x = in(1); // Value being mod'd | |
| 1028 Node *divisor = in(2); // Also is mask | |
| 0 | 1029 |
| 145 | 1030 hook->init_req(0, x); // Add a use to x to prevent him from dying |
| 1031 // Generate code to reduce X rapidly to nearly 2^k-1. | |
| 1032 for( int i = 0; i < trip_count; i++ ) { | |
| 0 | 1033 Node *xl = phase->transform( new (phase->C, 3) AndLNode(x,divisor) ); |
| 1034 Node *xh = phase->transform( new (phase->C, 3) RShiftLNode(x,phase->intcon(k)) ); // Must be signed | |
| 1035 x = phase->transform( new (phase->C, 3) AddLNode(xh,xl) ); | |
| 1036 hook->set_req(0, x); // Add a use to x to prevent him from dying | |
| 145 | 1037 } |
| 1038 | |
| 1039 // Generate sign-fixup code. Was original value positive? | |
| 1040 // long hack_res = (i >= 0) ? divisor : CONST64(1); | |
| 1041 Node *cmp1 = phase->transform( new (phase->C, 3) CmpLNode( in(1), phase->longcon(0) ) ); | |
| 1042 Node *bol1 = phase->transform( new (phase->C, 2) BoolNode( cmp1, BoolTest::ge ) ); | |
| 1043 Node *cmov1= phase->transform( new (phase->C, 4) CMoveLNode(bol1, phase->longcon(1), divisor, TypeLong::LONG) ); | |
| 1044 // if( x >= hack_res ) x -= divisor; | |
| 1045 Node *sub = phase->transform( new (phase->C, 3) SubLNode( x, divisor ) ); | |
| 1046 Node *cmp2 = phase->transform( new (phase->C, 3) CmpLNode( x, cmov1 ) ); | |
| 1047 Node *bol2 = phase->transform( new (phase->C, 2) BoolNode( cmp2, BoolTest::ge ) ); | |
| 1048 // Convention is to not transform the return value of an Ideal | |
| 1049 // since Ideal is expected to return a modified 'this' or a new node. | |
| 1050 Node *cmov2= new (phase->C, 4) CMoveLNode(bol2, x, sub, TypeLong::LONG); | |
| 1051 // cmov2 is now the mod | |
| 1052 | |
| 1053 // Now remove the bogus extra edges used to keep things alive | |
| 1054 if (can_reshape) { | |
| 1055 phase->is_IterGVN()->remove_dead_node(hook); | |
| 1056 } else { | |
| 1057 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
| 1058 } | |
| 1059 return cmov2; | |
| 0 | 1060 } |
| 145 | 1061 } |
| 1062 | |
| 1063 // Fell thru, the unroll case is not appropriate. Transform the modulo | |
| 1064 // into a long multiply/int multiply/subtract case | |
| 1065 | |
|
1914
ae065c367d93
6987135: Performance regression on Intel platform with 32-bits edition between 6u13 and 6u14.
kvn
parents:
1552
diff
changeset
|
1066 // Cannot handle mod 0, and min_jlong isn't handled by the transform |
| 145 | 1067 if( con == 0 || con == min_jlong ) return NULL; |
| 1068 | |
| 1069 // Get the absolute value of the constant; at this point, we can use this | |
| 1070 jlong pos_con = (con >= 0) ? con : -con; | |
| 1071 | |
| 1072 // integer Mod 1 is always 0 | |
| 1073 if( pos_con == 1 ) return new (phase->C, 1) ConLNode(TypeLong::ZERO); | |
| 1074 | |
| 1075 int log2_con = -1; | |
| 1076 | |
| 605 | 1077 // If this is a power of two, then maybe we can mask it |
| 145 | 1078 if( is_power_of_2_long(pos_con) ) { |
|
1914
ae065c367d93
6987135: Performance regression on Intel platform with 32-bits edition between 6u13 and 6u14.
kvn
parents:
1552
diff
changeset
|
1079 log2_con = exact_log2_long(pos_con); |
| 145 | 1080 |
| 1081 const Type *dt = phase->type(in(1)); | |
| 1082 const TypeLong *dtl = dt->isa_long(); | |
| 1083 | |
| 1084 // See if this can be masked, if the dividend is non-negative | |
| 1085 if( dtl && dtl->_lo >= 0 ) | |
| 1086 return ( new (phase->C, 3) AndLNode( in(1), phase->longcon( pos_con-1 ) ) ); | |
| 1087 } | |
| 0 | 1088 |
| 145 | 1089 // Save in(1) so that it cannot be changed or deleted |
| 1090 hook->init_req(0, in(1)); | |
| 1091 | |
|
1914
ae065c367d93
6987135: Performance regression on Intel platform with 32-bits edition between 6u13 and 6u14.
kvn
parents:
1552
diff
changeset
|
1092 // Divide using the transform from DivL to MulL |
| 145 | 1093 Node *result = transform_long_divide( phase, in(1), pos_con ); |
| 1094 if (result != NULL) { | |
| 1095 Node *divide = phase->transform(result); | |
| 1096 | |
| 1097 // Re-multiply, using a shift if this is a power of two | |
| 1098 Node *mult = NULL; | |
| 1099 | |
| 1100 if( log2_con >= 0 ) | |
| 1101 mult = phase->transform( new (phase->C, 3) LShiftLNode( divide, phase->intcon( log2_con ) ) ); | |
| 1102 else | |
| 1103 mult = phase->transform( new (phase->C, 3) MulLNode( divide, phase->longcon( pos_con ) ) ); | |
| 1104 | |
| 1105 // Finally, subtract the multiplied divided value from the original | |
| 1106 result = new (phase->C, 3) SubLNode( in(1), mult ); | |
| 0 | 1107 } |
| 145 | 1108 |
| 1109 // Now remove the bogus extra edges used to keep things alive | |
| 1110 if (can_reshape) { | |
| 1111 phase->is_IterGVN()->remove_dead_node(hook); | |
| 1112 } else { | |
| 1113 hook->set_req(0, NULL); // Just yank bogus edge during Parse phase | |
| 1114 } | |
| 1115 | |
| 1116 // return the value | |
| 1117 return result; | |
| 0 | 1118 } |
| 1119 | |
| 1120 //------------------------------Value------------------------------------------ | |
| 1121 const Type *ModLNode::Value( PhaseTransform *phase ) const { | |
| 1122 // Either input is TOP ==> the result is TOP | |
| 1123 const Type *t1 = phase->type( in(1) ); | |
| 1124 const Type *t2 = phase->type( in(2) ); | |
| 1125 if( t1 == Type::TOP ) return Type::TOP; | |
| 1126 if( t2 == Type::TOP ) return Type::TOP; | |
| 1127 | |
| 1128 // We always generate the dynamic check for 0. | |
| 1129 // 0 MOD X is 0 | |
| 1130 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; | |
| 1131 // X MOD X is 0 | |
| 1132 if( phase->eqv( in(1), in(2) ) ) return TypeLong::ZERO; | |
| 1133 | |
| 1134 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 1135 const Type *bot = bottom_type(); | |
| 1136 if( (t1 == bot) || (t2 == bot) || | |
| 1137 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 1138 return bot; | |
| 1139 | |
| 1140 const TypeLong *i1 = t1->is_long(); | |
| 1141 const TypeLong *i2 = t2->is_long(); | |
| 1142 if( !i1->is_con() || !i2->is_con() ) { | |
| 1143 if( i1->_lo >= CONST64(0) && i2->_lo >= CONST64(0) ) | |
| 1144 return TypeLong::POS; | |
| 1145 // If both numbers are not constants, we know little. | |
| 1146 return TypeLong::LONG; | |
| 1147 } | |
| 1148 // Mod by zero? Throw exception at runtime! | |
| 1149 if( !i2->get_con() ) return TypeLong::POS; | |
| 1150 | |
| 1151 // We must be modulo'ing 2 float constants. | |
| 1152 // Check for min_jint % '-1', result is defined to be '0'. | |
| 1153 if( i1->get_con() == min_jlong && i2->get_con() == -1 ) | |
| 1154 return TypeLong::ZERO; | |
| 1155 | |
| 1156 return TypeLong::make( i1->get_con() % i2->get_con() ); | |
| 1157 } | |
| 1158 | |
| 1159 | |
| 1160 //============================================================================= | |
| 1161 //------------------------------Value------------------------------------------ | |
| 1162 const Type *ModFNode::Value( PhaseTransform *phase ) const { | |
| 1163 // Either input is TOP ==> the result is TOP | |
| 1164 const Type *t1 = phase->type( in(1) ); | |
| 1165 const Type *t2 = phase->type( in(2) ); | |
| 1166 if( t1 == Type::TOP ) return Type::TOP; | |
| 1167 if( t2 == Type::TOP ) return Type::TOP; | |
| 1168 | |
| 1169 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 1170 const Type *bot = bottom_type(); | |
| 1171 if( (t1 == bot) || (t2 == bot) || | |
| 1172 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 1173 return bot; | |
| 1174 | |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1175 // 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
|
1176 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
|
1177 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
|
1178 } |
| 0 | 1179 |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1180 float f1 = t1->getf(); |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1181 float f2 = t2->getf(); |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1182 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
|
1183 jint x2 = jint_cast(f2); |
| 0 | 1184 |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1185 // 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
|
1186 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
|
1187 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
|
1188 |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1189 // 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
|
1190 if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jint) |
| 0 | 1191 return Type::FLOAT; |
| 1192 | |
| 1193 // We must be modulo'ing 2 float constants. | |
| 1194 // 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
|
1195 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
|
1196 if ((x1 ^ xr) < 0) { |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1197 xr ^= min_jint; |
| 0 | 1198 } |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1199 |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1200 return TypeF::make(jfloat_cast(xr)); |
| 0 | 1201 } |
| 1202 | |
| 1203 | |
| 1204 //============================================================================= | |
| 1205 //------------------------------Value------------------------------------------ | |
| 1206 const Type *ModDNode::Value( PhaseTransform *phase ) const { | |
| 1207 // Either input is TOP ==> the result is TOP | |
| 1208 const Type *t1 = phase->type( in(1) ); | |
| 1209 const Type *t2 = phase->type( in(2) ); | |
| 1210 if( t1 == Type::TOP ) return Type::TOP; | |
| 1211 if( t2 == Type::TOP ) return Type::TOP; | |
| 1212 | |
| 1213 // Either input is BOTTOM ==> the result is the local BOTTOM | |
| 1214 const Type *bot = bottom_type(); | |
| 1215 if( (t1 == bot) || (t2 == bot) || | |
| 1216 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
| 1217 return bot; | |
| 1218 | |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1219 // 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
|
1220 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
|
1221 return Type::DOUBLE; // note: x%x can be either NaN or 0 |
| 0 | 1222 } |
| 1223 | |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1224 double f1 = t1->getd(); |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1225 double f2 = t2->getd(); |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1226 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
|
1227 jlong x2 = jlong_cast(f2); |
| 0 | 1228 |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1229 // 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
|
1230 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
|
1231 if (g_isnan(f2)) return t2; |
| 0 | 1232 |
|
131
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1233 // 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
|
1234 if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jlong) |
| 0 | 1235 return Type::DOUBLE; |
| 1236 | |
| 1237 // 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
|
1238 // 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
|
1239 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
|
1240 if ((x1 ^ xr) < 0) { |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1241 xr ^= min_jlong; |
|
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 |
|
6e825ad773c6
6695288: runThese tests expr30303 and drem00301m1 fail when compiled code executes without deopt
jrose
parents:
0
diff
changeset
|
1244 return TypeD::make(jdouble_cast(xr)); |
| 0 | 1245 } |
| 1246 | |
| 1247 //============================================================================= | |
| 1248 | |
| 1249 DivModNode::DivModNode( Node *c, Node *dividend, Node *divisor ) : MultiNode(3) { | |
| 1250 init_req(0, c); | |
| 1251 init_req(1, dividend); | |
| 1252 init_req(2, divisor); | |
| 1253 } | |
| 1254 | |
| 1255 //------------------------------make------------------------------------------ | |
| 1256 DivModINode* DivModINode::make(Compile* C, Node* div_or_mod) { | |
| 1257 Node* n = div_or_mod; | |
| 1258 assert(n->Opcode() == Op_DivI || n->Opcode() == Op_ModI, | |
| 1259 "only div or mod input pattern accepted"); | |
| 1260 | |
| 1261 DivModINode* divmod = new (C, 3) DivModINode(n->in(0), n->in(1), n->in(2)); | |
| 1262 Node* dproj = new (C, 1) ProjNode(divmod, DivModNode::div_proj_num); | |
| 1263 Node* mproj = new (C, 1) ProjNode(divmod, DivModNode::mod_proj_num); | |
| 1264 return divmod; | |
| 1265 } | |
| 1266 | |
| 1267 //------------------------------make------------------------------------------ | |
| 1268 DivModLNode* DivModLNode::make(Compile* C, Node* div_or_mod) { | |
| 1269 Node* n = div_or_mod; | |
| 1270 assert(n->Opcode() == Op_DivL || n->Opcode() == Op_ModL, | |
| 1271 "only div or mod input pattern accepted"); | |
| 1272 | |
| 1273 DivModLNode* divmod = new (C, 3) DivModLNode(n->in(0), n->in(1), n->in(2)); | |
| 1274 Node* dproj = new (C, 1) ProjNode(divmod, DivModNode::div_proj_num); | |
| 1275 Node* mproj = new (C, 1) ProjNode(divmod, DivModNode::mod_proj_num); | |
| 1276 return divmod; | |
| 1277 } | |
| 1278 | |
| 1279 //------------------------------match------------------------------------------ | |
| 1280 // return result(s) along with their RegMask info | |
| 1281 Node *DivModINode::match( const ProjNode *proj, const Matcher *match ) { | |
| 1282 uint ideal_reg = proj->ideal_reg(); | |
| 1283 RegMask rm; | |
| 1284 if (proj->_con == div_proj_num) { | |
| 1285 rm = match->divI_proj_mask(); | |
| 1286 } else { | |
| 1287 assert(proj->_con == mod_proj_num, "must be div or mod projection"); | |
| 1288 rm = match->modI_proj_mask(); | |
| 1289 } | |
| 1290 return new (match->C, 1)MachProjNode(this, proj->_con, rm, ideal_reg); | |
| 1291 } | |
| 1292 | |
| 1293 | |
| 1294 //------------------------------match------------------------------------------ | |
| 1295 // return result(s) along with their RegMask info | |
| 1296 Node *DivModLNode::match( const ProjNode *proj, const Matcher *match ) { | |
| 1297 uint ideal_reg = proj->ideal_reg(); | |
| 1298 RegMask rm; | |
| 1299 if (proj->_con == div_proj_num) { | |
| 1300 rm = match->divL_proj_mask(); | |
| 1301 } else { | |
| 1302 assert(proj->_con == mod_proj_num, "must be div or mod projection"); | |
| 1303 rm = match->modL_proj_mask(); | |
| 1304 } | |
| 1305 return new (match->C, 1)MachProjNode(this, proj->_con, rm, ideal_reg); | |
| 1306 } |