Tue, 07 May 2019 20:38:26 +0000
8176100: [REDO][REDO] G1 Needs pre barrier on dereference of weak JNI handles
Summary: Add tag bit to all JNI weak handles
Reviewed-by: kbarrett, coleenp, tschatzl
goetz@6458 | 1 | /* |
phh@9669 | 2 | * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved. |
phh@9669 | 3 | * Copyright (c) 2012, 2017, SAP SE. All rights reserved. |
goetz@6458 | 4 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
goetz@6458 | 5 | * |
goetz@6458 | 6 | * This code is free software; you can redistribute it and/or modify it |
goetz@6458 | 7 | * under the terms of the GNU General Public License version 2 only, as |
goetz@6458 | 8 | * published by the Free Software Foundation. |
goetz@6458 | 9 | * |
goetz@6458 | 10 | * This code is distributed in the hope that it will be useful, but WITHOUT |
goetz@6458 | 11 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
goetz@6458 | 12 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
goetz@6458 | 13 | * version 2 for more details (a copy is included in the LICENSE file that |
goetz@6458 | 14 | * accompanied this code). |
goetz@6458 | 15 | * |
goetz@6458 | 16 | * You should have received a copy of the GNU General Public License version |
goetz@6458 | 17 | * 2 along with this work; if not, write to the Free Software Foundation, |
goetz@6458 | 18 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
goetz@6458 | 19 | * |
goetz@6458 | 20 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
goetz@6458 | 21 | * or visit www.oracle.com if you need additional information or have any |
goetz@6458 | 22 | * questions. |
goetz@6458 | 23 | * |
goetz@6458 | 24 | */ |
goetz@6458 | 25 | |
goetz@6458 | 26 | #include "precompiled.hpp" |
goetz@6458 | 27 | #include "asm/macroAssembler.inline.hpp" |
goetz@6458 | 28 | #include "compiler/disassembler.hpp" |
goetz@6458 | 29 | #include "gc_interface/collectedHeap.inline.hpp" |
goetz@6458 | 30 | #include "interpreter/interpreter.hpp" |
goetz@6458 | 31 | #include "memory/cardTableModRefBS.hpp" |
goetz@6458 | 32 | #include "memory/resourceArea.hpp" |
goetz@6458 | 33 | #include "prims/methodHandles.hpp" |
goetz@6458 | 34 | #include "runtime/biasedLocking.hpp" |
goetz@6458 | 35 | #include "runtime/interfaceSupport.hpp" |
goetz@6458 | 36 | #include "runtime/objectMonitor.hpp" |
goetz@6458 | 37 | #include "runtime/os.hpp" |
goetz@6458 | 38 | #include "runtime/sharedRuntime.hpp" |
goetz@6458 | 39 | #include "runtime/stubRoutines.hpp" |
goetz@6458 | 40 | #include "utilities/macros.hpp" |
goetz@6458 | 41 | #if INCLUDE_ALL_GCS |
goetz@6458 | 42 | #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" |
goetz@6458 | 43 | #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp" |
goetz@6458 | 44 | #include "gc_implementation/g1/heapRegion.hpp" |
goetz@6458 | 45 | #endif // INCLUDE_ALL_GCS |
goetz@6458 | 46 | |
goetz@6458 | 47 | #ifdef PRODUCT |
goetz@6458 | 48 | #define BLOCK_COMMENT(str) // nothing |
goetz@6458 | 49 | #else |
goetz@6458 | 50 | #define BLOCK_COMMENT(str) block_comment(str) |
goetz@6458 | 51 | #endif |
gromero@9496 | 52 | #define BIND(label) bind(label); BLOCK_COMMENT(#label ":") |
goetz@6458 | 53 | |
goetz@6458 | 54 | #ifdef ASSERT |
goetz@6458 | 55 | // On RISC, there's no benefit to verifying instruction boundaries. |
goetz@6458 | 56 | bool AbstractAssembler::pd_check_instruction_mark() { return false; } |
goetz@6458 | 57 | #endif |
goetz@6458 | 58 | |
goetz@6458 | 59 | void MacroAssembler::ld_largeoffset_unchecked(Register d, int si31, Register a, int emit_filler_nop) { |
goetz@6458 | 60 | assert(Assembler::is_simm(si31, 31) && si31 >= 0, "si31 out of range"); |
goetz@6458 | 61 | if (Assembler::is_simm(si31, 16)) { |
goetz@6458 | 62 | ld(d, si31, a); |
goetz@6458 | 63 | if (emit_filler_nop) nop(); |
goetz@6458 | 64 | } else { |
goetz@6458 | 65 | const int hi = MacroAssembler::largeoffset_si16_si16_hi(si31); |
goetz@6458 | 66 | const int lo = MacroAssembler::largeoffset_si16_si16_lo(si31); |
goetz@6458 | 67 | addis(d, a, hi); |
goetz@6458 | 68 | ld(d, lo, d); |
goetz@6458 | 69 | } |
goetz@6458 | 70 | } |
goetz@6458 | 71 | |
goetz@6458 | 72 | void MacroAssembler::ld_largeoffset(Register d, int si31, Register a, int emit_filler_nop) { |
goetz@6458 | 73 | assert_different_registers(d, a); |
goetz@6458 | 74 | ld_largeoffset_unchecked(d, si31, a, emit_filler_nop); |
goetz@6458 | 75 | } |
goetz@6458 | 76 | |
goetz@6458 | 77 | void MacroAssembler::load_sized_value(Register dst, RegisterOrConstant offs, Register base, |
goetz@6458 | 78 | size_t size_in_bytes, bool is_signed) { |
goetz@6458 | 79 | switch (size_in_bytes) { |
goetz@6458 | 80 | case 8: ld(dst, offs, base); break; |
goetz@6458 | 81 | case 4: is_signed ? lwa(dst, offs, base) : lwz(dst, offs, base); break; |
goetz@6458 | 82 | case 2: is_signed ? lha(dst, offs, base) : lhz(dst, offs, base); break; |
goetz@6458 | 83 | case 1: lbz(dst, offs, base); if (is_signed) extsb(dst, dst); break; // lba doesn't exist :( |
goetz@6458 | 84 | default: ShouldNotReachHere(); |
goetz@6458 | 85 | } |
goetz@6458 | 86 | } |
goetz@6458 | 87 | |
goetz@6458 | 88 | void MacroAssembler::store_sized_value(Register dst, RegisterOrConstant offs, Register base, |
goetz@6458 | 89 | size_t size_in_bytes) { |
goetz@6458 | 90 | switch (size_in_bytes) { |
goetz@6458 | 91 | case 8: std(dst, offs, base); break; |
goetz@6458 | 92 | case 4: stw(dst, offs, base); break; |
goetz@6458 | 93 | case 2: sth(dst, offs, base); break; |
goetz@6458 | 94 | case 1: stb(dst, offs, base); break; |
goetz@6458 | 95 | default: ShouldNotReachHere(); |
goetz@6458 | 96 | } |
goetz@6458 | 97 | } |
goetz@6458 | 98 | |
goetz@6495 | 99 | void MacroAssembler::align(int modulus, int max, int rem) { |
goetz@6495 | 100 | int padding = (rem + modulus - (offset() % modulus)) % modulus; |
goetz@6495 | 101 | if (padding > max) return; |
goetz@6495 | 102 | for (int c = (padding >> 2); c > 0; --c) { nop(); } |
goetz@6458 | 103 | } |
goetz@6458 | 104 | |
goetz@6458 | 105 | // Issue instructions that calculate given TOC from global TOC. |
goetz@6458 | 106 | void MacroAssembler::calculate_address_from_global_toc(Register dst, address addr, bool hi16, bool lo16, |
goetz@6458 | 107 | bool add_relocation, bool emit_dummy_addr) { |
goetz@6458 | 108 | int offset = -1; |
goetz@6458 | 109 | if (emit_dummy_addr) { |
goetz@6458 | 110 | offset = -128; // dummy address |
goetz@6458 | 111 | } else if (addr != (address)(intptr_t)-1) { |
goetz@6458 | 112 | offset = MacroAssembler::offset_to_global_toc(addr); |
goetz@6458 | 113 | } |
goetz@6458 | 114 | |
goetz@6458 | 115 | if (hi16) { |
goetz@6458 | 116 | addis(dst, R29, MacroAssembler::largeoffset_si16_si16_hi(offset)); |
goetz@6458 | 117 | } |
goetz@6458 | 118 | if (lo16) { |
goetz@6458 | 119 | if (add_relocation) { |
goetz@6458 | 120 | // Relocate at the addi to avoid confusion with a load from the method's TOC. |
goetz@6458 | 121 | relocate(internal_word_Relocation::spec(addr)); |
goetz@6458 | 122 | } |
goetz@6458 | 123 | addi(dst, dst, MacroAssembler::largeoffset_si16_si16_lo(offset)); |
goetz@6458 | 124 | } |
goetz@6458 | 125 | } |
goetz@6458 | 126 | |
goetz@6458 | 127 | int MacroAssembler::patch_calculate_address_from_global_toc_at(address a, address bound, address addr) { |
goetz@6458 | 128 | const int offset = MacroAssembler::offset_to_global_toc(addr); |
goetz@6458 | 129 | |
goetz@6458 | 130 | const address inst2_addr = a; |
goetz@6458 | 131 | const int inst2 = *(int *)inst2_addr; |
goetz@6458 | 132 | |
goetz@6458 | 133 | // The relocation points to the second instruction, the addi, |
goetz@6458 | 134 | // and the addi reads and writes the same register dst. |
goetz@6458 | 135 | const int dst = inv_rt_field(inst2); |
goetz@6458 | 136 | assert(is_addi(inst2) && inv_ra_field(inst2) == dst, "must be addi reading and writing dst"); |
goetz@6458 | 137 | |
goetz@6458 | 138 | // Now, find the preceding addis which writes to dst. |
goetz@6458 | 139 | int inst1 = 0; |
goetz@6458 | 140 | address inst1_addr = inst2_addr - BytesPerInstWord; |
goetz@6458 | 141 | while (inst1_addr >= bound) { |
goetz@6458 | 142 | inst1 = *(int *) inst1_addr; |
goetz@6458 | 143 | if (is_addis(inst1) && inv_rt_field(inst1) == dst) { |
goetz@6458 | 144 | // Stop, found the addis which writes dst. |
goetz@6458 | 145 | break; |
goetz@6458 | 146 | } |
goetz@6458 | 147 | inst1_addr -= BytesPerInstWord; |
goetz@6458 | 148 | } |
goetz@6458 | 149 | |
goetz@6458 | 150 | assert(is_addis(inst1) && inv_ra_field(inst1) == 29 /* R29 */, "source must be global TOC"); |
goetz@6458 | 151 | set_imm((int *)inst1_addr, MacroAssembler::largeoffset_si16_si16_hi(offset)); |
goetz@6458 | 152 | set_imm((int *)inst2_addr, MacroAssembler::largeoffset_si16_si16_lo(offset)); |
goetz@6458 | 153 | return (int)((intptr_t)addr - (intptr_t)inst1_addr); |
goetz@6458 | 154 | } |
goetz@6458 | 155 | |
goetz@6458 | 156 | address MacroAssembler::get_address_of_calculate_address_from_global_toc_at(address a, address bound) { |
goetz@6458 | 157 | const address inst2_addr = a; |
goetz@6458 | 158 | const int inst2 = *(int *)inst2_addr; |
goetz@6458 | 159 | |
goetz@6458 | 160 | // The relocation points to the second instruction, the addi, |
goetz@6458 | 161 | // and the addi reads and writes the same register dst. |
goetz@6458 | 162 | const int dst = inv_rt_field(inst2); |
goetz@6458 | 163 | assert(is_addi(inst2) && inv_ra_field(inst2) == dst, "must be addi reading and writing dst"); |
goetz@6458 | 164 | |
goetz@6458 | 165 | // Now, find the preceding addis which writes to dst. |
goetz@6458 | 166 | int inst1 = 0; |
goetz@6458 | 167 | address inst1_addr = inst2_addr - BytesPerInstWord; |
goetz@6458 | 168 | while (inst1_addr >= bound) { |
goetz@6458 | 169 | inst1 = *(int *) inst1_addr; |
goetz@6458 | 170 | if (is_addis(inst1) && inv_rt_field(inst1) == dst) { |
goetz@6458 | 171 | // stop, found the addis which writes dst |
goetz@6458 | 172 | break; |
goetz@6458 | 173 | } |
goetz@6458 | 174 | inst1_addr -= BytesPerInstWord; |
goetz@6458 | 175 | } |
goetz@6458 | 176 | |
goetz@6458 | 177 | assert(is_addis(inst1) && inv_ra_field(inst1) == 29 /* R29 */, "source must be global TOC"); |
goetz@6458 | 178 | |
goetz@6458 | 179 | int offset = (get_imm(inst1_addr, 0) << 16) + get_imm(inst2_addr, 0); |
goetz@6458 | 180 | // -1 is a special case |
goetz@6458 | 181 | if (offset == -1) { |
goetz@6458 | 182 | return (address)(intptr_t)-1; |
goetz@6458 | 183 | } else { |
goetz@6458 | 184 | return global_toc() + offset; |
goetz@6458 | 185 | } |
goetz@6458 | 186 | } |
goetz@6458 | 187 | |
goetz@6458 | 188 | #ifdef _LP64 |
goetz@6458 | 189 | // Patch compressed oops or klass constants. |
goetz@6495 | 190 | // Assembler sequence is |
goetz@6495 | 191 | // 1) compressed oops: |
goetz@6495 | 192 | // lis rx = const.hi |
goetz@6495 | 193 | // ori rx = rx | const.lo |
goetz@6495 | 194 | // 2) compressed klass: |
goetz@6495 | 195 | // lis rx = const.hi |
goetz@6495 | 196 | // clrldi rx = rx & 0xFFFFffff // clearMS32b, optional |
goetz@6495 | 197 | // ori rx = rx | const.lo |
goetz@6495 | 198 | // Clrldi will be passed by. |
goetz@6458 | 199 | int MacroAssembler::patch_set_narrow_oop(address a, address bound, narrowOop data) { |
goetz@6458 | 200 | assert(UseCompressedOops, "Should only patch compressed oops"); |
goetz@6458 | 201 | |
goetz@6458 | 202 | const address inst2_addr = a; |
goetz@6458 | 203 | const int inst2 = *(int *)inst2_addr; |
goetz@6458 | 204 | |
goetz@6495 | 205 | // The relocation points to the second instruction, the ori, |
goetz@6495 | 206 | // and the ori reads and writes the same register dst. |
goetz@6495 | 207 | const int dst = inv_rta_field(inst2); |
goetz@6501 | 208 | assert(is_ori(inst2) && inv_rs_field(inst2) == dst, "must be ori reading and writing dst"); |
goetz@6458 | 209 | // Now, find the preceding addis which writes to dst. |
goetz@6458 | 210 | int inst1 = 0; |
goetz@6458 | 211 | address inst1_addr = inst2_addr - BytesPerInstWord; |
goetz@6458 | 212 | bool inst1_found = false; |
goetz@6458 | 213 | while (inst1_addr >= bound) { |
goetz@6458 | 214 | inst1 = *(int *)inst1_addr; |
goetz@6458 | 215 | if (is_lis(inst1) && inv_rs_field(inst1) == dst) { inst1_found = true; break; } |
goetz@6458 | 216 | inst1_addr -= BytesPerInstWord; |
goetz@6458 | 217 | } |
goetz@6458 | 218 | assert(inst1_found, "inst is not lis"); |
goetz@6458 | 219 | |
goetz@6458 | 220 | int xc = (data >> 16) & 0xffff; |
goetz@6458 | 221 | int xd = (data >> 0) & 0xffff; |
goetz@6458 | 222 | |
goetz@6495 | 223 | set_imm((int *)inst1_addr, (short)(xc)); // see enc_load_con_narrow_hi/_lo |
goetz@6501 | 224 | set_imm((int *)inst2_addr, (xd)); // unsigned int |
goetz@6458 | 225 | return (int)((intptr_t)inst2_addr - (intptr_t)inst1_addr); |
goetz@6458 | 226 | } |
goetz@6458 | 227 | |
goetz@6458 | 228 | // Get compressed oop or klass constant. |
goetz@6458 | 229 | narrowOop MacroAssembler::get_narrow_oop(address a, address bound) { |
goetz@6458 | 230 | assert(UseCompressedOops, "Should only patch compressed oops"); |
goetz@6458 | 231 | |
goetz@6458 | 232 | const address inst2_addr = a; |
goetz@6458 | 233 | const int inst2 = *(int *)inst2_addr; |
goetz@6458 | 234 | |
goetz@6495 | 235 | // The relocation points to the second instruction, the ori, |
goetz@6495 | 236 | // and the ori reads and writes the same register dst. |
goetz@6495 | 237 | const int dst = inv_rta_field(inst2); |
goetz@6501 | 238 | assert(is_ori(inst2) && inv_rs_field(inst2) == dst, "must be ori reading and writing dst"); |
goetz@6458 | 239 | // Now, find the preceding lis which writes to dst. |
goetz@6458 | 240 | int inst1 = 0; |
goetz@6458 | 241 | address inst1_addr = inst2_addr - BytesPerInstWord; |
goetz@6458 | 242 | bool inst1_found = false; |
goetz@6458 | 243 | |
goetz@6458 | 244 | while (inst1_addr >= bound) { |
goetz@6458 | 245 | inst1 = *(int *) inst1_addr; |
goetz@6458 | 246 | if (is_lis(inst1) && inv_rs_field(inst1) == dst) { inst1_found = true; break;} |
goetz@6458 | 247 | inst1_addr -= BytesPerInstWord; |
goetz@6458 | 248 | } |
goetz@6458 | 249 | assert(inst1_found, "inst is not lis"); |
goetz@6458 | 250 | |
goetz@6495 | 251 | uint xl = ((unsigned int) (get_imm(inst2_addr, 0) & 0xffff)); |
goetz@6495 | 252 | uint xh = (((get_imm(inst1_addr, 0)) & 0xffff) << 16); |
goetz@6495 | 253 | |
goetz@6458 | 254 | return (int) (xl | xh); |
goetz@6458 | 255 | } |
goetz@6458 | 256 | #endif // _LP64 |
goetz@6458 | 257 | |
goetz@6458 | 258 | void MacroAssembler::load_const_from_method_toc(Register dst, AddressLiteral& a, Register toc) { |
goetz@6458 | 259 | int toc_offset = 0; |
goetz@6458 | 260 | // Use RelocationHolder::none for the constant pool entry, otherwise |
goetz@6458 | 261 | // we will end up with a failing NativeCall::verify(x) where x is |
goetz@6458 | 262 | // the address of the constant pool entry. |
goetz@6458 | 263 | // FIXME: We should insert relocation information for oops at the constant |
goetz@6458 | 264 | // pool entries instead of inserting it at the loads; patching of a constant |
goetz@6458 | 265 | // pool entry should be less expensive. |
goetz@6495 | 266 | address oop_address = address_constant((address)a.value(), RelocationHolder::none); |
goetz@6495 | 267 | // Relocate at the pc of the load. |
goetz@6495 | 268 | relocate(a.rspec()); |
goetz@6495 | 269 | toc_offset = (int)(oop_address - code()->consts()->start()); |
goetz@6458 | 270 | ld_largeoffset_unchecked(dst, toc_offset, toc, true); |
goetz@6458 | 271 | } |
goetz@6458 | 272 | |
goetz@6458 | 273 | bool MacroAssembler::is_load_const_from_method_toc_at(address a) { |
goetz@6458 | 274 | const address inst1_addr = a; |
goetz@6458 | 275 | const int inst1 = *(int *)inst1_addr; |
goetz@6458 | 276 | |
goetz@6458 | 277 | // The relocation points to the ld or the addis. |
goetz@6458 | 278 | return (is_ld(inst1)) || |
goetz@6458 | 279 | (is_addis(inst1) && inv_ra_field(inst1) != 0); |
goetz@6458 | 280 | } |
goetz@6458 | 281 | |
goetz@6458 | 282 | int MacroAssembler::get_offset_of_load_const_from_method_toc_at(address a) { |
goetz@6458 | 283 | assert(is_load_const_from_method_toc_at(a), "must be load_const_from_method_toc"); |
goetz@6458 | 284 | |
goetz@6458 | 285 | const address inst1_addr = a; |
goetz@6458 | 286 | const int inst1 = *(int *)inst1_addr; |
goetz@6458 | 287 | |
goetz@6458 | 288 | if (is_ld(inst1)) { |
goetz@6458 | 289 | return inv_d1_field(inst1); |
goetz@6458 | 290 | } else if (is_addis(inst1)) { |
goetz@6458 | 291 | const int dst = inv_rt_field(inst1); |
goetz@6458 | 292 | |
goetz@6458 | 293 | // Now, find the succeeding ld which reads and writes to dst. |
goetz@6458 | 294 | address inst2_addr = inst1_addr + BytesPerInstWord; |
goetz@6458 | 295 | int inst2 = 0; |
goetz@6458 | 296 | while (true) { |
goetz@6458 | 297 | inst2 = *(int *) inst2_addr; |
goetz@6458 | 298 | if (is_ld(inst2) && inv_ra_field(inst2) == dst && inv_rt_field(inst2) == dst) { |
goetz@6458 | 299 | // Stop, found the ld which reads and writes dst. |
goetz@6458 | 300 | break; |
goetz@6458 | 301 | } |
goetz@6458 | 302 | inst2_addr += BytesPerInstWord; |
goetz@6458 | 303 | } |
goetz@6458 | 304 | return (inv_d1_field(inst1) << 16) + inv_d1_field(inst2); |
goetz@6458 | 305 | } |
goetz@6458 | 306 | ShouldNotReachHere(); |
goetz@6458 | 307 | return 0; |
goetz@6458 | 308 | } |
goetz@6458 | 309 | |
goetz@6458 | 310 | // Get the constant from a `load_const' sequence. |
goetz@6458 | 311 | long MacroAssembler::get_const(address a) { |
goetz@6458 | 312 | assert(is_load_const_at(a), "not a load of a constant"); |
goetz@6458 | 313 | const int *p = (const int*) a; |
goetz@6458 | 314 | unsigned long x = (((unsigned long) (get_imm(a,0) & 0xffff)) << 48); |
goetz@6458 | 315 | if (is_ori(*(p+1))) { |
goetz@6458 | 316 | x |= (((unsigned long) (get_imm(a,1) & 0xffff)) << 32); |
goetz@6458 | 317 | x |= (((unsigned long) (get_imm(a,3) & 0xffff)) << 16); |
goetz@6458 | 318 | x |= (((unsigned long) (get_imm(a,4) & 0xffff))); |
goetz@6458 | 319 | } else if (is_lis(*(p+1))) { |
goetz@6458 | 320 | x |= (((unsigned long) (get_imm(a,2) & 0xffff)) << 32); |
goetz@6458 | 321 | x |= (((unsigned long) (get_imm(a,1) & 0xffff)) << 16); |
goetz@6458 | 322 | x |= (((unsigned long) (get_imm(a,3) & 0xffff))); |
goetz@6458 | 323 | } else { |
goetz@6458 | 324 | ShouldNotReachHere(); |
goetz@6458 | 325 | return (long) 0; |
goetz@6458 | 326 | } |
goetz@6458 | 327 | return (long) x; |
goetz@6458 | 328 | } |
goetz@6458 | 329 | |
goetz@6458 | 330 | // Patch the 64 bit constant of a `load_const' sequence. This is a low |
goetz@6458 | 331 | // level procedure. It neither flushes the instruction cache nor is it |
goetz@6458 | 332 | // mt safe. |
goetz@6458 | 333 | void MacroAssembler::patch_const(address a, long x) { |
goetz@6458 | 334 | assert(is_load_const_at(a), "not a load of a constant"); |
goetz@6458 | 335 | int *p = (int*) a; |
goetz@6458 | 336 | if (is_ori(*(p+1))) { |
goetz@6458 | 337 | set_imm(0 + p, (x >> 48) & 0xffff); |
goetz@6458 | 338 | set_imm(1 + p, (x >> 32) & 0xffff); |
goetz@6458 | 339 | set_imm(3 + p, (x >> 16) & 0xffff); |
goetz@6458 | 340 | set_imm(4 + p, x & 0xffff); |
goetz@6458 | 341 | } else if (is_lis(*(p+1))) { |
goetz@6458 | 342 | set_imm(0 + p, (x >> 48) & 0xffff); |
goetz@6458 | 343 | set_imm(2 + p, (x >> 32) & 0xffff); |
goetz@6458 | 344 | set_imm(1 + p, (x >> 16) & 0xffff); |
goetz@6458 | 345 | set_imm(3 + p, x & 0xffff); |
goetz@6458 | 346 | } else { |
goetz@6458 | 347 | ShouldNotReachHere(); |
goetz@6458 | 348 | } |
goetz@6458 | 349 | } |
goetz@6458 | 350 | |
goetz@6458 | 351 | AddressLiteral MacroAssembler::allocate_metadata_address(Metadata* obj) { |
goetz@6458 | 352 | assert(oop_recorder() != NULL, "this assembler needs a Recorder"); |
goetz@6458 | 353 | int index = oop_recorder()->allocate_metadata_index(obj); |
goetz@6458 | 354 | RelocationHolder rspec = metadata_Relocation::spec(index); |
goetz@6458 | 355 | return AddressLiteral((address)obj, rspec); |
goetz@6458 | 356 | } |
goetz@6458 | 357 | |
goetz@6458 | 358 | AddressLiteral MacroAssembler::constant_metadata_address(Metadata* obj) { |
goetz@6458 | 359 | assert(oop_recorder() != NULL, "this assembler needs a Recorder"); |
goetz@6458 | 360 | int index = oop_recorder()->find_index(obj); |
goetz@6458 | 361 | RelocationHolder rspec = metadata_Relocation::spec(index); |
goetz@6458 | 362 | return AddressLiteral((address)obj, rspec); |
goetz@6458 | 363 | } |
goetz@6458 | 364 | |
goetz@6458 | 365 | AddressLiteral MacroAssembler::allocate_oop_address(jobject obj) { |
goetz@6458 | 366 | assert(oop_recorder() != NULL, "this assembler needs an OopRecorder"); |
goetz@6458 | 367 | int oop_index = oop_recorder()->allocate_oop_index(obj); |
goetz@6458 | 368 | return AddressLiteral(address(obj), oop_Relocation::spec(oop_index)); |
goetz@6458 | 369 | } |
goetz@6458 | 370 | |
goetz@6458 | 371 | AddressLiteral MacroAssembler::constant_oop_address(jobject obj) { |
goetz@6458 | 372 | assert(oop_recorder() != NULL, "this assembler needs an OopRecorder"); |
goetz@6458 | 373 | int oop_index = oop_recorder()->find_index(obj); |
goetz@6458 | 374 | return AddressLiteral(address(obj), oop_Relocation::spec(oop_index)); |
goetz@6458 | 375 | } |
goetz@6458 | 376 | |
goetz@6458 | 377 | RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr, |
goetz@6458 | 378 | Register tmp, int offset) { |
goetz@6458 | 379 | intptr_t value = *delayed_value_addr; |
goetz@6458 | 380 | if (value != 0) { |
goetz@6458 | 381 | return RegisterOrConstant(value + offset); |
goetz@6458 | 382 | } |
goetz@6458 | 383 | |
goetz@6458 | 384 | // Load indirectly to solve generation ordering problem. |
goetz@6458 | 385 | // static address, no relocation |
goetz@6458 | 386 | int simm16_offset = load_const_optimized(tmp, delayed_value_addr, noreg, true); |
goetz@6458 | 387 | ld(tmp, simm16_offset, tmp); // must be aligned ((xa & 3) == 0) |
goetz@6458 | 388 | |
goetz@6458 | 389 | if (offset != 0) { |
goetz@6458 | 390 | addi(tmp, tmp, offset); |
goetz@6458 | 391 | } |
goetz@6458 | 392 | |
goetz@6458 | 393 | return RegisterOrConstant(tmp); |
goetz@6458 | 394 | } |
goetz@6458 | 395 | |
goetz@6458 | 396 | #ifndef PRODUCT |
goetz@6458 | 397 | void MacroAssembler::pd_print_patched_instruction(address branch) { |
goetz@6458 | 398 | Unimplemented(); // TODO: PPC port |
goetz@6458 | 399 | } |
goetz@6458 | 400 | #endif // ndef PRODUCT |
goetz@6458 | 401 | |
goetz@6458 | 402 | // Conditional far branch for destinations encodable in 24+2 bits. |
goetz@6458 | 403 | void MacroAssembler::bc_far(int boint, int biint, Label& dest, int optimize) { |
goetz@6458 | 404 | |
goetz@6458 | 405 | // If requested by flag optimize, relocate the bc_far as a |
goetz@6458 | 406 | // runtime_call and prepare for optimizing it when the code gets |
goetz@6458 | 407 | // relocated. |
goetz@6458 | 408 | if (optimize == bc_far_optimize_on_relocate) { |
goetz@6458 | 409 | relocate(relocInfo::runtime_call_type); |
goetz@6458 | 410 | } |
goetz@6458 | 411 | |
goetz@6458 | 412 | // variant 2: |
goetz@6458 | 413 | // |
goetz@6458 | 414 | // b!cxx SKIP |
goetz@6458 | 415 | // bxx DEST |
goetz@6458 | 416 | // SKIP: |
goetz@6458 | 417 | // |
goetz@6458 | 418 | |
goetz@6458 | 419 | const int opposite_boint = add_bhint_to_boint(opposite_bhint(inv_boint_bhint(boint)), |
goetz@6458 | 420 | opposite_bcond(inv_boint_bcond(boint))); |
goetz@6458 | 421 | |
goetz@6458 | 422 | // We emit two branches. |
goetz@6458 | 423 | // First, a conditional branch which jumps around the far branch. |
goetz@6458 | 424 | const address not_taken_pc = pc() + 2 * BytesPerInstWord; |
goetz@6458 | 425 | const address bc_pc = pc(); |
goetz@6458 | 426 | bc(opposite_boint, biint, not_taken_pc); |
goetz@6458 | 427 | |
goetz@6458 | 428 | const int bc_instr = *(int*)bc_pc; |
goetz@6458 | 429 | assert(not_taken_pc == (address)inv_bd_field(bc_instr, (intptr_t)bc_pc), "postcondition"); |
goetz@6458 | 430 | assert(opposite_boint == inv_bo_field(bc_instr), "postcondition"); |
goetz@6458 | 431 | assert(boint == add_bhint_to_boint(opposite_bhint(inv_boint_bhint(inv_bo_field(bc_instr))), |
goetz@6458 | 432 | opposite_bcond(inv_boint_bcond(inv_bo_field(bc_instr)))), |
goetz@6458 | 433 | "postcondition"); |
goetz@6458 | 434 | assert(biint == inv_bi_field(bc_instr), "postcondition"); |
goetz@6458 | 435 | |
goetz@6458 | 436 | // Second, an unconditional far branch which jumps to dest. |
goetz@6458 | 437 | // Note: target(dest) remembers the current pc (see CodeSection::target) |
goetz@6458 | 438 | // and returns the current pc if the label is not bound yet; when |
goetz@6458 | 439 | // the label gets bound, the unconditional far branch will be patched. |
goetz@6458 | 440 | const address target_pc = target(dest); |
goetz@6458 | 441 | const address b_pc = pc(); |
goetz@6458 | 442 | b(target_pc); |
goetz@6458 | 443 | |
goetz@6458 | 444 | assert(not_taken_pc == pc(), "postcondition"); |
goetz@6458 | 445 | assert(dest.is_bound() || target_pc == b_pc, "postcondition"); |
goetz@6458 | 446 | } |
goetz@6458 | 447 | |
goetz@6458 | 448 | bool MacroAssembler::is_bc_far_at(address instruction_addr) { |
goetz@6458 | 449 | return is_bc_far_variant1_at(instruction_addr) || |
goetz@6458 | 450 | is_bc_far_variant2_at(instruction_addr) || |
goetz@6458 | 451 | is_bc_far_variant3_at(instruction_addr); |
goetz@6458 | 452 | } |
goetz@6458 | 453 | |
goetz@6458 | 454 | address MacroAssembler::get_dest_of_bc_far_at(address instruction_addr) { |
goetz@6458 | 455 | if (is_bc_far_variant1_at(instruction_addr)) { |
goetz@6458 | 456 | const address instruction_1_addr = instruction_addr; |
goetz@6458 | 457 | const int instruction_1 = *(int*)instruction_1_addr; |
goetz@6458 | 458 | return (address)inv_bd_field(instruction_1, (intptr_t)instruction_1_addr); |
goetz@6458 | 459 | } else if (is_bc_far_variant2_at(instruction_addr)) { |
goetz@6458 | 460 | const address instruction_2_addr = instruction_addr + 4; |
goetz@6458 | 461 | return bxx_destination(instruction_2_addr); |
goetz@6458 | 462 | } else if (is_bc_far_variant3_at(instruction_addr)) { |
goetz@6458 | 463 | return instruction_addr + 8; |
goetz@6458 | 464 | } |
goetz@6458 | 465 | // variant 4 ??? |
goetz@6458 | 466 | ShouldNotReachHere(); |
goetz@6458 | 467 | return NULL; |
goetz@6458 | 468 | } |
goetz@6458 | 469 | void MacroAssembler::set_dest_of_bc_far_at(address instruction_addr, address dest) { |
goetz@6458 | 470 | |
goetz@6458 | 471 | if (is_bc_far_variant3_at(instruction_addr)) { |
goetz@6458 | 472 | // variant 3, far cond branch to the next instruction, already patched to nops: |
goetz@6458 | 473 | // |
goetz@6458 | 474 | // nop |
goetz@6458 | 475 | // endgroup |
goetz@6458 | 476 | // SKIP/DEST: |
goetz@6458 | 477 | // |
goetz@6458 | 478 | return; |
goetz@6458 | 479 | } |
goetz@6458 | 480 | |
goetz@6458 | 481 | // first, extract boint and biint from the current branch |
goetz@6458 | 482 | int boint = 0; |
goetz@6458 | 483 | int biint = 0; |
goetz@6458 | 484 | |
goetz@6458 | 485 | ResourceMark rm; |
goetz@6458 | 486 | const int code_size = 2 * BytesPerInstWord; |
goetz@6458 | 487 | CodeBuffer buf(instruction_addr, code_size); |
goetz@6458 | 488 | MacroAssembler masm(&buf); |
goetz@6458 | 489 | if (is_bc_far_variant2_at(instruction_addr) && dest == instruction_addr + 8) { |
goetz@6458 | 490 | // Far branch to next instruction: Optimize it by patching nops (produce variant 3). |
goetz@6458 | 491 | masm.nop(); |
goetz@6458 | 492 | masm.endgroup(); |
goetz@6458 | 493 | } else { |
goetz@6458 | 494 | if (is_bc_far_variant1_at(instruction_addr)) { |
goetz@6458 | 495 | // variant 1, the 1st instruction contains the destination address: |
goetz@6458 | 496 | // |
goetz@6458 | 497 | // bcxx DEST |
goetz@6458 | 498 | // endgroup |
goetz@6458 | 499 | // |
goetz@6458 | 500 | const int instruction_1 = *(int*)(instruction_addr); |
goetz@6458 | 501 | boint = inv_bo_field(instruction_1); |
goetz@6458 | 502 | biint = inv_bi_field(instruction_1); |
goetz@6458 | 503 | } else if (is_bc_far_variant2_at(instruction_addr)) { |
goetz@6458 | 504 | // variant 2, the 2nd instruction contains the destination address: |
goetz@6458 | 505 | // |
goetz@6458 | 506 | // b!cxx SKIP |
goetz@6458 | 507 | // bxx DEST |
goetz@6458 | 508 | // SKIP: |
goetz@6458 | 509 | // |
goetz@6458 | 510 | const int instruction_1 = *(int*)(instruction_addr); |
goetz@6458 | 511 | boint = add_bhint_to_boint(opposite_bhint(inv_boint_bhint(inv_bo_field(instruction_1))), |
goetz@6458 | 512 | opposite_bcond(inv_boint_bcond(inv_bo_field(instruction_1)))); |
goetz@6458 | 513 | biint = inv_bi_field(instruction_1); |
goetz@6458 | 514 | } else { |
goetz@6458 | 515 | // variant 4??? |
goetz@6458 | 516 | ShouldNotReachHere(); |
goetz@6458 | 517 | } |
goetz@6458 | 518 | |
goetz@6458 | 519 | // second, set the new branch destination and optimize the code |
goetz@6458 | 520 | if (dest != instruction_addr + 4 && // the bc_far is still unbound! |
goetz@6458 | 521 | masm.is_within_range_of_bcxx(dest, instruction_addr)) { |
goetz@6458 | 522 | // variant 1: |
goetz@6458 | 523 | // |
goetz@6458 | 524 | // bcxx DEST |
goetz@6458 | 525 | // endgroup |
goetz@6458 | 526 | // |
goetz@6458 | 527 | masm.bc(boint, biint, dest); |
goetz@6458 | 528 | masm.endgroup(); |
goetz@6458 | 529 | } else { |
goetz@6458 | 530 | // variant 2: |
goetz@6458 | 531 | // |
goetz@6458 | 532 | // b!cxx SKIP |
goetz@6458 | 533 | // bxx DEST |
goetz@6458 | 534 | // SKIP: |
goetz@6458 | 535 | // |
goetz@6458 | 536 | const int opposite_boint = add_bhint_to_boint(opposite_bhint(inv_boint_bhint(boint)), |
goetz@6458 | 537 | opposite_bcond(inv_boint_bcond(boint))); |
goetz@6458 | 538 | const address not_taken_pc = masm.pc() + 2 * BytesPerInstWord; |
goetz@6458 | 539 | masm.bc(opposite_boint, biint, not_taken_pc); |
goetz@6458 | 540 | masm.b(dest); |
goetz@6458 | 541 | } |
goetz@6458 | 542 | } |
goetz@6495 | 543 | ICache::ppc64_flush_icache_bytes(instruction_addr, code_size); |
goetz@6458 | 544 | } |
goetz@6458 | 545 | |
goetz@6458 | 546 | // Emit a NOT mt-safe patchable 64 bit absolute call/jump. |
goetz@6458 | 547 | void MacroAssembler::bxx64_patchable(address dest, relocInfo::relocType rt, bool link) { |
goetz@6458 | 548 | // get current pc |
goetz@6458 | 549 | uint64_t start_pc = (uint64_t) pc(); |
goetz@6458 | 550 | |
goetz@6458 | 551 | const address pc_of_bl = (address) (start_pc + (6*BytesPerInstWord)); // bl is last |
goetz@6458 | 552 | const address pc_of_b = (address) (start_pc + (0*BytesPerInstWord)); // b is first |
goetz@6458 | 553 | |
goetz@6458 | 554 | // relocate here |
goetz@6458 | 555 | if (rt != relocInfo::none) { |
goetz@6458 | 556 | relocate(rt); |
goetz@6458 | 557 | } |
goetz@6458 | 558 | |
goetz@6458 | 559 | if ( ReoptimizeCallSequences && |
goetz@6458 | 560 | (( link && is_within_range_of_b(dest, pc_of_bl)) || |
goetz@6458 | 561 | (!link && is_within_range_of_b(dest, pc_of_b)))) { |
goetz@6458 | 562 | // variant 2: |
goetz@6458 | 563 | // Emit an optimized, pc-relative call/jump. |
goetz@6458 | 564 | |
goetz@6458 | 565 | if (link) { |
goetz@6458 | 566 | // some padding |
goetz@6458 | 567 | nop(); |
goetz@6458 | 568 | nop(); |
goetz@6458 | 569 | nop(); |
goetz@6458 | 570 | nop(); |
goetz@6458 | 571 | nop(); |
goetz@6458 | 572 | nop(); |
goetz@6458 | 573 | |
goetz@6458 | 574 | // do the call |
goetz@6458 | 575 | assert(pc() == pc_of_bl, "just checking"); |
goetz@6458 | 576 | bl(dest, relocInfo::none); |
goetz@6458 | 577 | } else { |
goetz@6458 | 578 | // do the jump |
goetz@6458 | 579 | assert(pc() == pc_of_b, "just checking"); |
goetz@6458 | 580 | b(dest, relocInfo::none); |
goetz@6458 | 581 | |
goetz@6458 | 582 | // some padding |
goetz@6458 | 583 | nop(); |
goetz@6458 | 584 | nop(); |
goetz@6458 | 585 | nop(); |
goetz@6458 | 586 | nop(); |
goetz@6458 | 587 | nop(); |
goetz@6458 | 588 | nop(); |
goetz@6458 | 589 | } |
goetz@6458 | 590 | |
goetz@6458 | 591 | // Assert that we can identify the emitted call/jump. |
goetz@6458 | 592 | assert(is_bxx64_patchable_variant2_at((address)start_pc, link), |
goetz@6458 | 593 | "can't identify emitted call"); |
goetz@6458 | 594 | } else { |
goetz@6458 | 595 | // variant 1: |
goetz@6458 | 596 | mr(R0, R11); // spill R11 -> R0. |
goetz@6458 | 597 | |
goetz@6458 | 598 | // Load the destination address into CTR, |
goetz@6458 | 599 | // calculate destination relative to global toc. |
goetz@6458 | 600 | calculate_address_from_global_toc(R11, dest, true, true, false); |
goetz@6458 | 601 | |
goetz@6458 | 602 | mtctr(R11); |
goetz@6458 | 603 | mr(R11, R0); // spill R11 <- R0. |
goetz@6458 | 604 | nop(); |
goetz@6458 | 605 | |
goetz@6458 | 606 | // do the call/jump |
goetz@6458 | 607 | if (link) { |
goetz@6458 | 608 | bctrl(); |
goetz@6458 | 609 | } else{ |
goetz@6458 | 610 | bctr(); |
goetz@6458 | 611 | } |
goetz@6458 | 612 | // Assert that we can identify the emitted call/jump. |
goetz@6458 | 613 | assert(is_bxx64_patchable_variant1b_at((address)start_pc, link), |
goetz@6458 | 614 | "can't identify emitted call"); |
goetz@6458 | 615 | } |
goetz@6458 | 616 | |
goetz@6458 | 617 | // Assert that we can identify the emitted call/jump. |
goetz@6458 | 618 | assert(is_bxx64_patchable_at((address)start_pc, link), |
goetz@6458 | 619 | "can't identify emitted call"); |
goetz@6458 | 620 | assert(get_dest_of_bxx64_patchable_at((address)start_pc, link) == dest, |
goetz@6458 | 621 | "wrong encoding of dest address"); |
goetz@6458 | 622 | } |
goetz@6458 | 623 | |
goetz@6458 | 624 | // Identify a bxx64_patchable instruction. |
goetz@6458 | 625 | bool MacroAssembler::is_bxx64_patchable_at(address instruction_addr, bool link) { |
goetz@6458 | 626 | return is_bxx64_patchable_variant1b_at(instruction_addr, link) |
goetz@6458 | 627 | //|| is_bxx64_patchable_variant1_at(instruction_addr, link) |
goetz@6458 | 628 | || is_bxx64_patchable_variant2_at(instruction_addr, link); |
goetz@6458 | 629 | } |
goetz@6458 | 630 | |
goetz@6458 | 631 | // Does the call64_patchable instruction use a pc-relative encoding of |
goetz@6458 | 632 | // the call destination? |
goetz@6458 | 633 | bool MacroAssembler::is_bxx64_patchable_pcrelative_at(address instruction_addr, bool link) { |
goetz@6458 | 634 | // variant 2 is pc-relative |
goetz@6458 | 635 | return is_bxx64_patchable_variant2_at(instruction_addr, link); |
goetz@6458 | 636 | } |
goetz@6458 | 637 | |
goetz@6458 | 638 | // Identify variant 1. |
goetz@6458 | 639 | bool MacroAssembler::is_bxx64_patchable_variant1_at(address instruction_addr, bool link) { |
goetz@6458 | 640 | unsigned int* instr = (unsigned int*) instruction_addr; |
goetz@6458 | 641 | return (link ? is_bctrl(instr[6]) : is_bctr(instr[6])) // bctr[l] |
goetz@6458 | 642 | && is_mtctr(instr[5]) // mtctr |
goetz@6458 | 643 | && is_load_const_at(instruction_addr); |
goetz@6458 | 644 | } |
goetz@6458 | 645 | |
goetz@6458 | 646 | // Identify variant 1b: load destination relative to global toc. |
goetz@6458 | 647 | bool MacroAssembler::is_bxx64_patchable_variant1b_at(address instruction_addr, bool link) { |
goetz@6458 | 648 | unsigned int* instr = (unsigned int*) instruction_addr; |
goetz@6458 | 649 | return (link ? is_bctrl(instr[6]) : is_bctr(instr[6])) // bctr[l] |
goetz@6458 | 650 | && is_mtctr(instr[3]) // mtctr |
goetz@6458 | 651 | && is_calculate_address_from_global_toc_at(instruction_addr + 2*BytesPerInstWord, instruction_addr); |
goetz@6458 | 652 | } |
goetz@6458 | 653 | |
goetz@6458 | 654 | // Identify variant 2. |
goetz@6458 | 655 | bool MacroAssembler::is_bxx64_patchable_variant2_at(address instruction_addr, bool link) { |
goetz@6458 | 656 | unsigned int* instr = (unsigned int*) instruction_addr; |
goetz@6458 | 657 | if (link) { |
goetz@6458 | 658 | return is_bl (instr[6]) // bl dest is last |
goetz@6458 | 659 | && is_nop(instr[0]) // nop |
goetz@6458 | 660 | && is_nop(instr[1]) // nop |
goetz@6458 | 661 | && is_nop(instr[2]) // nop |
goetz@6458 | 662 | && is_nop(instr[3]) // nop |
goetz@6458 | 663 | && is_nop(instr[4]) // nop |
goetz@6458 | 664 | && is_nop(instr[5]); // nop |
goetz@6458 | 665 | } else { |
goetz@6458 | 666 | return is_b (instr[0]) // b dest is first |
goetz@6458 | 667 | && is_nop(instr[1]) // nop |
goetz@6458 | 668 | && is_nop(instr[2]) // nop |
goetz@6458 | 669 | && is_nop(instr[3]) // nop |
goetz@6458 | 670 | && is_nop(instr[4]) // nop |
goetz@6458 | 671 | && is_nop(instr[5]) // nop |
goetz@6458 | 672 | && is_nop(instr[6]); // nop |
goetz@6458 | 673 | } |
goetz@6458 | 674 | } |
goetz@6458 | 675 | |
goetz@6458 | 676 | // Set dest address of a bxx64_patchable instruction. |
goetz@6458 | 677 | void MacroAssembler::set_dest_of_bxx64_patchable_at(address instruction_addr, address dest, bool link) { |
goetz@6458 | 678 | ResourceMark rm; |
goetz@6458 | 679 | int code_size = MacroAssembler::bxx64_patchable_size; |
goetz@6458 | 680 | CodeBuffer buf(instruction_addr, code_size); |
goetz@6458 | 681 | MacroAssembler masm(&buf); |
goetz@6458 | 682 | masm.bxx64_patchable(dest, relocInfo::none, link); |
goetz@6495 | 683 | ICache::ppc64_flush_icache_bytes(instruction_addr, code_size); |
goetz@6458 | 684 | } |
goetz@6458 | 685 | |
goetz@6458 | 686 | // Get dest address of a bxx64_patchable instruction. |
goetz@6458 | 687 | address MacroAssembler::get_dest_of_bxx64_patchable_at(address instruction_addr, bool link) { |
goetz@6458 | 688 | if (is_bxx64_patchable_variant1_at(instruction_addr, link)) { |
goetz@6458 | 689 | return (address) (unsigned long) get_const(instruction_addr); |
goetz@6458 | 690 | } else if (is_bxx64_patchable_variant2_at(instruction_addr, link)) { |
goetz@6458 | 691 | unsigned int* instr = (unsigned int*) instruction_addr; |
goetz@6458 | 692 | if (link) { |
goetz@6458 | 693 | const int instr_idx = 6; // bl is last |
goetz@6458 | 694 | int branchoffset = branch_destination(instr[instr_idx], 0); |
goetz@6458 | 695 | return instruction_addr + branchoffset + instr_idx*BytesPerInstWord; |
goetz@6458 | 696 | } else { |
goetz@6458 | 697 | const int instr_idx = 0; // b is first |
goetz@6458 | 698 | int branchoffset = branch_destination(instr[instr_idx], 0); |
goetz@6458 | 699 | return instruction_addr + branchoffset + instr_idx*BytesPerInstWord; |
goetz@6458 | 700 | } |
goetz@6458 | 701 | // Load dest relative to global toc. |
goetz@6458 | 702 | } else if (is_bxx64_patchable_variant1b_at(instruction_addr, link)) { |
goetz@6458 | 703 | return get_address_of_calculate_address_from_global_toc_at(instruction_addr + 2*BytesPerInstWord, |
goetz@6458 | 704 | instruction_addr); |
goetz@6458 | 705 | } else { |
goetz@6458 | 706 | ShouldNotReachHere(); |
goetz@6458 | 707 | return NULL; |
goetz@6458 | 708 | } |
goetz@6458 | 709 | } |
goetz@6458 | 710 | |
goetz@6458 | 711 | // Uses ordering which corresponds to ABI: |
goetz@6458 | 712 | // _savegpr0_14: std r14,-144(r1) |
goetz@6458 | 713 | // _savegpr0_15: std r15,-136(r1) |
goetz@6458 | 714 | // _savegpr0_16: std r16,-128(r1) |
goetz@6458 | 715 | void MacroAssembler::save_nonvolatile_gprs(Register dst, int offset) { |
goetz@6458 | 716 | std(R14, offset, dst); offset += 8; |
goetz@6458 | 717 | std(R15, offset, dst); offset += 8; |
goetz@6458 | 718 | std(R16, offset, dst); offset += 8; |
goetz@6458 | 719 | std(R17, offset, dst); offset += 8; |
goetz@6458 | 720 | std(R18, offset, dst); offset += 8; |
goetz@6458 | 721 | std(R19, offset, dst); offset += 8; |
goetz@6458 | 722 | std(R20, offset, dst); offset += 8; |
goetz@6458 | 723 | std(R21, offset, dst); offset += 8; |
goetz@6458 | 724 | std(R22, offset, dst); offset += 8; |
goetz@6458 | 725 | std(R23, offset, dst); offset += 8; |
goetz@6458 | 726 | std(R24, offset, dst); offset += 8; |
goetz@6458 | 727 | std(R25, offset, dst); offset += 8; |
goetz@6458 | 728 | std(R26, offset, dst); offset += 8; |
goetz@6458 | 729 | std(R27, offset, dst); offset += 8; |
goetz@6458 | 730 | std(R28, offset, dst); offset += 8; |
goetz@6458 | 731 | std(R29, offset, dst); offset += 8; |
goetz@6458 | 732 | std(R30, offset, dst); offset += 8; |
goetz@6458 | 733 | std(R31, offset, dst); offset += 8; |
goetz@6458 | 734 | |
goetz@6458 | 735 | stfd(F14, offset, dst); offset += 8; |
goetz@6458 | 736 | stfd(F15, offset, dst); offset += 8; |
goetz@6458 | 737 | stfd(F16, offset, dst); offset += 8; |
goetz@6458 | 738 | stfd(F17, offset, dst); offset += 8; |
goetz@6458 | 739 | stfd(F18, offset, dst); offset += 8; |
goetz@6458 | 740 | stfd(F19, offset, dst); offset += 8; |
goetz@6458 | 741 | stfd(F20, offset, dst); offset += 8; |
goetz@6458 | 742 | stfd(F21, offset, dst); offset += 8; |
goetz@6458 | 743 | stfd(F22, offset, dst); offset += 8; |
goetz@6458 | 744 | stfd(F23, offset, dst); offset += 8; |
goetz@6458 | 745 | stfd(F24, offset, dst); offset += 8; |
goetz@6458 | 746 | stfd(F25, offset, dst); offset += 8; |
goetz@6458 | 747 | stfd(F26, offset, dst); offset += 8; |
goetz@6458 | 748 | stfd(F27, offset, dst); offset += 8; |
goetz@6458 | 749 | stfd(F28, offset, dst); offset += 8; |
goetz@6458 | 750 | stfd(F29, offset, dst); offset += 8; |
goetz@6458 | 751 | stfd(F30, offset, dst); offset += 8; |
goetz@6458 | 752 | stfd(F31, offset, dst); |
goetz@6458 | 753 | } |
goetz@6458 | 754 | |
goetz@6458 | 755 | // Uses ordering which corresponds to ABI: |
goetz@6458 | 756 | // _restgpr0_14: ld r14,-144(r1) |
goetz@6458 | 757 | // _restgpr0_15: ld r15,-136(r1) |
goetz@6458 | 758 | // _restgpr0_16: ld r16,-128(r1) |
goetz@6458 | 759 | void MacroAssembler::restore_nonvolatile_gprs(Register src, int offset) { |
goetz@6458 | 760 | ld(R14, offset, src); offset += 8; |
goetz@6458 | 761 | ld(R15, offset, src); offset += 8; |
goetz@6458 | 762 | ld(R16, offset, src); offset += 8; |
goetz@6458 | 763 | ld(R17, offset, src); offset += 8; |
goetz@6458 | 764 | ld(R18, offset, src); offset += 8; |
goetz@6458 | 765 | ld(R19, offset, src); offset += 8; |
goetz@6458 | 766 | ld(R20, offset, src); offset += 8; |
goetz@6458 | 767 | ld(R21, offset, src); offset += 8; |
goetz@6458 | 768 | ld(R22, offset, src); offset += 8; |
goetz@6458 | 769 | ld(R23, offset, src); offset += 8; |
goetz@6458 | 770 | ld(R24, offset, src); offset += 8; |
goetz@6458 | 771 | ld(R25, offset, src); offset += 8; |
goetz@6458 | 772 | ld(R26, offset, src); offset += 8; |
goetz@6458 | 773 | ld(R27, offset, src); offset += 8; |
goetz@6458 | 774 | ld(R28, offset, src); offset += 8; |
goetz@6458 | 775 | ld(R29, offset, src); offset += 8; |
goetz@6458 | 776 | ld(R30, offset, src); offset += 8; |
goetz@6458 | 777 | ld(R31, offset, src); offset += 8; |
goetz@6458 | 778 | |
goetz@6458 | 779 | // FP registers |
goetz@6458 | 780 | lfd(F14, offset, src); offset += 8; |
goetz@6458 | 781 | lfd(F15, offset, src); offset += 8; |
goetz@6458 | 782 | lfd(F16, offset, src); offset += 8; |
goetz@6458 | 783 | lfd(F17, offset, src); offset += 8; |
goetz@6458 | 784 | lfd(F18, offset, src); offset += 8; |
goetz@6458 | 785 | lfd(F19, offset, src); offset += 8; |
goetz@6458 | 786 | lfd(F20, offset, src); offset += 8; |
goetz@6458 | 787 | lfd(F21, offset, src); offset += 8; |
goetz@6458 | 788 | lfd(F22, offset, src); offset += 8; |
goetz@6458 | 789 | lfd(F23, offset, src); offset += 8; |
goetz@6458 | 790 | lfd(F24, offset, src); offset += 8; |
goetz@6458 | 791 | lfd(F25, offset, src); offset += 8; |
goetz@6458 | 792 | lfd(F26, offset, src); offset += 8; |
goetz@6458 | 793 | lfd(F27, offset, src); offset += 8; |
goetz@6458 | 794 | lfd(F28, offset, src); offset += 8; |
goetz@6458 | 795 | lfd(F29, offset, src); offset += 8; |
goetz@6458 | 796 | lfd(F30, offset, src); offset += 8; |
goetz@6458 | 797 | lfd(F31, offset, src); |
goetz@6458 | 798 | } |
goetz@6458 | 799 | |
goetz@6458 | 800 | // For verify_oops. |
goetz@6458 | 801 | void MacroAssembler::save_volatile_gprs(Register dst, int offset) { |
goetz@7424 | 802 | std(R2, offset, dst); offset += 8; |
goetz@6458 | 803 | std(R3, offset, dst); offset += 8; |
goetz@6458 | 804 | std(R4, offset, dst); offset += 8; |
goetz@6458 | 805 | std(R5, offset, dst); offset += 8; |
goetz@6458 | 806 | std(R6, offset, dst); offset += 8; |
goetz@6458 | 807 | std(R7, offset, dst); offset += 8; |
goetz@6458 | 808 | std(R8, offset, dst); offset += 8; |
goetz@6458 | 809 | std(R9, offset, dst); offset += 8; |
goetz@6458 | 810 | std(R10, offset, dst); offset += 8; |
goetz@6458 | 811 | std(R11, offset, dst); offset += 8; |
goetz@6458 | 812 | std(R12, offset, dst); |
goetz@6458 | 813 | } |
goetz@6458 | 814 | |
goetz@6458 | 815 | // For verify_oops. |
goetz@6458 | 816 | void MacroAssembler::restore_volatile_gprs(Register src, int offset) { |
goetz@7424 | 817 | ld(R2, offset, src); offset += 8; |
goetz@6458 | 818 | ld(R3, offset, src); offset += 8; |
goetz@6458 | 819 | ld(R4, offset, src); offset += 8; |
goetz@6458 | 820 | ld(R5, offset, src); offset += 8; |
goetz@6458 | 821 | ld(R6, offset, src); offset += 8; |
goetz@6458 | 822 | ld(R7, offset, src); offset += 8; |
goetz@6458 | 823 | ld(R8, offset, src); offset += 8; |
goetz@6458 | 824 | ld(R9, offset, src); offset += 8; |
goetz@6458 | 825 | ld(R10, offset, src); offset += 8; |
goetz@6458 | 826 | ld(R11, offset, src); offset += 8; |
goetz@6458 | 827 | ld(R12, offset, src); |
goetz@6458 | 828 | } |
goetz@6458 | 829 | |
goetz@6458 | 830 | void MacroAssembler::save_LR_CR(Register tmp) { |
goetz@6458 | 831 | mfcr(tmp); |
goetz@6458 | 832 | std(tmp, _abi(cr), R1_SP); |
goetz@6458 | 833 | mflr(tmp); |
goetz@6458 | 834 | std(tmp, _abi(lr), R1_SP); |
goetz@6458 | 835 | // Tmp must contain lr on exit! (see return_addr and prolog in ppc64.ad) |
goetz@6458 | 836 | } |
goetz@6458 | 837 | |
goetz@6458 | 838 | void MacroAssembler::restore_LR_CR(Register tmp) { |
goetz@6458 | 839 | assert(tmp != R1_SP, "must be distinct"); |
goetz@6458 | 840 | ld(tmp, _abi(lr), R1_SP); |
goetz@6458 | 841 | mtlr(tmp); |
goetz@6458 | 842 | ld(tmp, _abi(cr), R1_SP); |
goetz@6458 | 843 | mtcr(tmp); |
goetz@6458 | 844 | } |
goetz@6458 | 845 | |
goetz@6458 | 846 | address MacroAssembler::get_PC_trash_LR(Register result) { |
goetz@6458 | 847 | Label L; |
goetz@6458 | 848 | bl(L); |
goetz@6458 | 849 | bind(L); |
goetz@6458 | 850 | address lr_pc = pc(); |
goetz@6458 | 851 | mflr(result); |
goetz@6458 | 852 | return lr_pc; |
goetz@6458 | 853 | } |
goetz@6458 | 854 | |
goetz@6458 | 855 | void MacroAssembler::resize_frame(Register offset, Register tmp) { |
goetz@6458 | 856 | #ifdef ASSERT |
goetz@6458 | 857 | assert_different_registers(offset, tmp, R1_SP); |
goetz@6458 | 858 | andi_(tmp, offset, frame::alignment_in_bytes-1); |
goetz@6458 | 859 | asm_assert_eq("resize_frame: unaligned", 0x204); |
goetz@6458 | 860 | #endif |
goetz@6458 | 861 | |
goetz@6458 | 862 | // tmp <- *(SP) |
goetz@6458 | 863 | ld(tmp, _abi(callers_sp), R1_SP); |
goetz@6458 | 864 | // addr <- SP + offset; |
goetz@6458 | 865 | // *(addr) <- tmp; |
goetz@6458 | 866 | // SP <- addr |
goetz@6458 | 867 | stdux(tmp, R1_SP, offset); |
goetz@6458 | 868 | } |
goetz@6458 | 869 | |
goetz@6458 | 870 | void MacroAssembler::resize_frame(int offset, Register tmp) { |
goetz@6458 | 871 | assert(is_simm(offset, 16), "too big an offset"); |
goetz@6458 | 872 | assert_different_registers(tmp, R1_SP); |
goetz@6458 | 873 | assert((offset & (frame::alignment_in_bytes-1))==0, "resize_frame: unaligned"); |
goetz@6458 | 874 | // tmp <- *(SP) |
goetz@6458 | 875 | ld(tmp, _abi(callers_sp), R1_SP); |
goetz@6458 | 876 | // addr <- SP + offset; |
goetz@6458 | 877 | // *(addr) <- tmp; |
goetz@6458 | 878 | // SP <- addr |
goetz@6458 | 879 | stdu(tmp, offset, R1_SP); |
goetz@6458 | 880 | } |
goetz@6458 | 881 | |
goetz@6458 | 882 | void MacroAssembler::resize_frame_absolute(Register addr, Register tmp1, Register tmp2) { |
goetz@6458 | 883 | // (addr == tmp1) || (addr == tmp2) is allowed here! |
goetz@6458 | 884 | assert(tmp1 != tmp2, "must be distinct"); |
goetz@6458 | 885 | |
goetz@6458 | 886 | // compute offset w.r.t. current stack pointer |
goetz@6458 | 887 | // tmp_1 <- addr - SP (!) |
goetz@6458 | 888 | subf(tmp1, R1_SP, addr); |
goetz@6458 | 889 | |
goetz@6458 | 890 | // atomically update SP keeping back link. |
goetz@6458 | 891 | resize_frame(tmp1/* offset */, tmp2/* tmp */); |
goetz@6458 | 892 | } |
goetz@6458 | 893 | |
goetz@6458 | 894 | void MacroAssembler::push_frame(Register bytes, Register tmp) { |
goetz@6458 | 895 | #ifdef ASSERT |
goetz@6458 | 896 | assert(bytes != R0, "r0 not allowed here"); |
goetz@6458 | 897 | andi_(R0, bytes, frame::alignment_in_bytes-1); |
goetz@6458 | 898 | asm_assert_eq("push_frame(Reg, Reg): unaligned", 0x203); |
goetz@6458 | 899 | #endif |
goetz@6458 | 900 | neg(tmp, bytes); |
goetz@6458 | 901 | stdux(R1_SP, R1_SP, tmp); |
goetz@6458 | 902 | } |
goetz@6458 | 903 | |
goetz@6458 | 904 | // Push a frame of size `bytes'. |
goetz@6458 | 905 | void MacroAssembler::push_frame(unsigned int bytes, Register tmp) { |
goetz@6458 | 906 | long offset = align_addr(bytes, frame::alignment_in_bytes); |
goetz@6458 | 907 | if (is_simm(-offset, 16)) { |
goetz@6458 | 908 | stdu(R1_SP, -offset, R1_SP); |
goetz@6458 | 909 | } else { |
goetz@6458 | 910 | load_const(tmp, -offset); |
goetz@6458 | 911 | stdux(R1_SP, R1_SP, tmp); |
goetz@6458 | 912 | } |
goetz@6458 | 913 | } |
goetz@6458 | 914 | |
goetz@6511 | 915 | // Push a frame of size `bytes' plus abi_reg_args on top. |
goetz@6511 | 916 | void MacroAssembler::push_frame_reg_args(unsigned int bytes, Register tmp) { |
goetz@6511 | 917 | push_frame(bytes + frame::abi_reg_args_size, tmp); |
goetz@6458 | 918 | } |
goetz@6458 | 919 | |
goetz@6458 | 920 | // Setup up a new C frame with a spill area for non-volatile GPRs and |
goetz@6458 | 921 | // additional space for local variables. |
goetz@6511 | 922 | void MacroAssembler::push_frame_reg_args_nonvolatiles(unsigned int bytes, |
goetz@6511 | 923 | Register tmp) { |
goetz@6511 | 924 | push_frame(bytes + frame::abi_reg_args_size + frame::spill_nonvolatiles_size, tmp); |
goetz@6458 | 925 | } |
goetz@6458 | 926 | |
goetz@6458 | 927 | // Pop current C frame. |
goetz@6458 | 928 | void MacroAssembler::pop_frame() { |
goetz@6458 | 929 | ld(R1_SP, _abi(callers_sp), R1_SP); |
goetz@6458 | 930 | } |
goetz@6458 | 931 | |
goetz@6511 | 932 | #if defined(ABI_ELFv2) |
goetz@6511 | 933 | address MacroAssembler::branch_to(Register r_function_entry, bool and_link) { |
goetz@6511 | 934 | // TODO(asmundak): make sure the caller uses R12 as function descriptor |
goetz@6511 | 935 | // most of the times. |
goetz@6511 | 936 | if (R12 != r_function_entry) { |
goetz@6511 | 937 | mr(R12, r_function_entry); |
goetz@6511 | 938 | } |
goetz@6511 | 939 | mtctr(R12); |
goetz@6511 | 940 | // Do a call or a branch. |
goetz@6511 | 941 | if (and_link) { |
goetz@6511 | 942 | bctrl(); |
goetz@6511 | 943 | } else { |
goetz@6511 | 944 | bctr(); |
goetz@6511 | 945 | } |
goetz@6511 | 946 | _last_calls_return_pc = pc(); |
goetz@6511 | 947 | |
goetz@6511 | 948 | return _last_calls_return_pc; |
goetz@6511 | 949 | } |
goetz@6511 | 950 | |
goetz@6511 | 951 | // Call a C function via a function descriptor and use full C |
goetz@6511 | 952 | // calling conventions. Updates and returns _last_calls_return_pc. |
goetz@6511 | 953 | address MacroAssembler::call_c(Register r_function_entry) { |
goetz@6511 | 954 | return branch_to(r_function_entry, /*and_link=*/true); |
goetz@6511 | 955 | } |
goetz@6511 | 956 | |
goetz@6511 | 957 | // For tail calls: only branch, don't link, so callee returns to caller of this function. |
goetz@6511 | 958 | address MacroAssembler::call_c_and_return_to_caller(Register r_function_entry) { |
goetz@6511 | 959 | return branch_to(r_function_entry, /*and_link=*/false); |
goetz@6511 | 960 | } |
goetz@6511 | 961 | |
goetz@6511 | 962 | address MacroAssembler::call_c(address function_entry, relocInfo::relocType rt) { |
goetz@6511 | 963 | load_const(R12, function_entry, R0); |
goetz@6511 | 964 | return branch_to(R12, /*and_link=*/true); |
goetz@6511 | 965 | } |
goetz@6511 | 966 | |
goetz@6511 | 967 | #else |
goetz@6458 | 968 | // Generic version of a call to C function via a function descriptor |
goetz@6458 | 969 | // with variable support for C calling conventions (TOC, ENV, etc.). |
goetz@6458 | 970 | // Updates and returns _last_calls_return_pc. |
goetz@6458 | 971 | address MacroAssembler::branch_to(Register function_descriptor, bool and_link, bool save_toc_before_call, |
goetz@6458 | 972 | bool restore_toc_after_call, bool load_toc_of_callee, bool load_env_of_callee) { |
goetz@6458 | 973 | // we emit standard ptrgl glue code here |
goetz@6458 | 974 | assert((function_descriptor != R0), "function_descriptor cannot be R0"); |
goetz@6458 | 975 | |
goetz@6458 | 976 | // retrieve necessary entries from the function descriptor |
goetz@6458 | 977 | ld(R0, in_bytes(FunctionDescriptor::entry_offset()), function_descriptor); |
goetz@6458 | 978 | mtctr(R0); |
goetz@6458 | 979 | |
goetz@6458 | 980 | if (load_toc_of_callee) { |
goetz@6458 | 981 | ld(R2_TOC, in_bytes(FunctionDescriptor::toc_offset()), function_descriptor); |
goetz@6458 | 982 | } |
goetz@6458 | 983 | if (load_env_of_callee) { |
goetz@6458 | 984 | ld(R11, in_bytes(FunctionDescriptor::env_offset()), function_descriptor); |
goetz@6458 | 985 | } else if (load_toc_of_callee) { |
goetz@6458 | 986 | li(R11, 0); |
goetz@6458 | 987 | } |
goetz@6458 | 988 | |
goetz@6458 | 989 | // do a call or a branch |
goetz@6458 | 990 | if (and_link) { |
goetz@6458 | 991 | bctrl(); |
goetz@6458 | 992 | } else { |
goetz@6458 | 993 | bctr(); |
goetz@6458 | 994 | } |
goetz@6458 | 995 | _last_calls_return_pc = pc(); |
goetz@6458 | 996 | |
goetz@6458 | 997 | return _last_calls_return_pc; |
goetz@6458 | 998 | } |
goetz@6458 | 999 | |
goetz@6458 | 1000 | // Call a C function via a function descriptor and use full C calling |
goetz@6458 | 1001 | // conventions. |
goetz@6458 | 1002 | // We don't use the TOC in generated code, so there is no need to save |
goetz@6458 | 1003 | // and restore its value. |
goetz@6458 | 1004 | address MacroAssembler::call_c(Register fd) { |
goetz@6458 | 1005 | return branch_to(fd, /*and_link=*/true, |
goetz@6458 | 1006 | /*save toc=*/false, |
goetz@6458 | 1007 | /*restore toc=*/false, |
goetz@6458 | 1008 | /*load toc=*/true, |
goetz@6458 | 1009 | /*load env=*/true); |
goetz@6458 | 1010 | } |
goetz@6458 | 1011 | |
goetz@6495 | 1012 | address MacroAssembler::call_c_and_return_to_caller(Register fd) { |
goetz@6495 | 1013 | return branch_to(fd, /*and_link=*/false, |
goetz@6495 | 1014 | /*save toc=*/false, |
goetz@6495 | 1015 | /*restore toc=*/false, |
goetz@6495 | 1016 | /*load toc=*/true, |
goetz@6495 | 1017 | /*load env=*/true); |
goetz@6495 | 1018 | } |
goetz@6495 | 1019 | |
goetz@6458 | 1020 | address MacroAssembler::call_c(const FunctionDescriptor* fd, relocInfo::relocType rt) { |
goetz@6458 | 1021 | if (rt != relocInfo::none) { |
goetz@6458 | 1022 | // this call needs to be relocatable |
goetz@6458 | 1023 | if (!ReoptimizeCallSequences |
goetz@6458 | 1024 | || (rt != relocInfo::runtime_call_type && rt != relocInfo::none) |
goetz@6458 | 1025 | || fd == NULL // support code-size estimation |
goetz@6458 | 1026 | || !fd->is_friend_function() |
goetz@6458 | 1027 | || fd->entry() == NULL) { |
goetz@6458 | 1028 | // it's not a friend function as defined by class FunctionDescriptor, |
goetz@6458 | 1029 | // so do a full call-c here. |
goetz@6458 | 1030 | load_const(R11, (address)fd, R0); |
goetz@6458 | 1031 | |
goetz@6458 | 1032 | bool has_env = (fd != NULL && fd->env() != NULL); |
goetz@6458 | 1033 | return branch_to(R11, /*and_link=*/true, |
goetz@6501 | 1034 | /*save toc=*/false, |
goetz@6501 | 1035 | /*restore toc=*/false, |
goetz@6501 | 1036 | /*load toc=*/true, |
goetz@6501 | 1037 | /*load env=*/has_env); |
goetz@6458 | 1038 | } else { |
goetz@6458 | 1039 | // It's a friend function. Load the entry point and don't care about |
goetz@6458 | 1040 | // toc and env. Use an optimizable call instruction, but ensure the |
goetz@6458 | 1041 | // same code-size as in the case of a non-friend function. |
goetz@6458 | 1042 | nop(); |
goetz@6458 | 1043 | nop(); |
goetz@6458 | 1044 | nop(); |
goetz@6458 | 1045 | bl64_patchable(fd->entry(), rt); |
goetz@6458 | 1046 | _last_calls_return_pc = pc(); |
goetz@6458 | 1047 | return _last_calls_return_pc; |
goetz@6458 | 1048 | } |
goetz@6458 | 1049 | } else { |
goetz@6458 | 1050 | // This call does not need to be relocatable, do more aggressive |
goetz@6458 | 1051 | // optimizations. |
goetz@6458 | 1052 | if (!ReoptimizeCallSequences |
goetz@6458 | 1053 | || !fd->is_friend_function()) { |
goetz@6458 | 1054 | // It's not a friend function as defined by class FunctionDescriptor, |
goetz@6458 | 1055 | // so do a full call-c here. |
goetz@6458 | 1056 | load_const(R11, (address)fd, R0); |
goetz@6458 | 1057 | return branch_to(R11, /*and_link=*/true, |
goetz@6501 | 1058 | /*save toc=*/false, |
goetz@6501 | 1059 | /*restore toc=*/false, |
goetz@6501 | 1060 | /*load toc=*/true, |
goetz@6501 | 1061 | /*load env=*/true); |
goetz@6458 | 1062 | } else { |
goetz@6458 | 1063 | // it's a friend function, load the entry point and don't care about |
goetz@6458 | 1064 | // toc and env. |
goetz@6458 | 1065 | address dest = fd->entry(); |
goetz@6458 | 1066 | if (is_within_range_of_b(dest, pc())) { |
goetz@6458 | 1067 | bl(dest); |
goetz@6458 | 1068 | } else { |
goetz@6458 | 1069 | bl64_patchable(dest, rt); |
goetz@6458 | 1070 | } |
goetz@6458 | 1071 | _last_calls_return_pc = pc(); |
goetz@6458 | 1072 | return _last_calls_return_pc; |
goetz@6458 | 1073 | } |
goetz@6458 | 1074 | } |
goetz@6458 | 1075 | } |
goetz@6458 | 1076 | |
goetz@6458 | 1077 | // Call a C function. All constants needed reside in TOC. |
goetz@6458 | 1078 | // |
goetz@6458 | 1079 | // Read the address to call from the TOC. |
goetz@6458 | 1080 | // Read env from TOC, if fd specifies an env. |
goetz@6458 | 1081 | // Read new TOC from TOC. |
goetz@6458 | 1082 | address MacroAssembler::call_c_using_toc(const FunctionDescriptor* fd, |
goetz@6458 | 1083 | relocInfo::relocType rt, Register toc) { |
goetz@6458 | 1084 | if (!ReoptimizeCallSequences |
goetz@6458 | 1085 | || (rt != relocInfo::runtime_call_type && rt != relocInfo::none) |
goetz@6458 | 1086 | || !fd->is_friend_function()) { |
goetz@6458 | 1087 | // It's not a friend function as defined by class FunctionDescriptor, |
goetz@6458 | 1088 | // so do a full call-c here. |
goetz@6458 | 1089 | assert(fd->entry() != NULL, "function must be linked"); |
goetz@6458 | 1090 | |
goetz@6458 | 1091 | AddressLiteral fd_entry(fd->entry()); |
goetz@6458 | 1092 | load_const_from_method_toc(R11, fd_entry, toc); |
goetz@6458 | 1093 | mtctr(R11); |
goetz@6458 | 1094 | if (fd->env() == NULL) { |
goetz@6458 | 1095 | li(R11, 0); |
goetz@6458 | 1096 | nop(); |
goetz@6458 | 1097 | } else { |
goetz@6458 | 1098 | AddressLiteral fd_env(fd->env()); |
goetz@6458 | 1099 | load_const_from_method_toc(R11, fd_env, toc); |
goetz@6458 | 1100 | } |
goetz@6458 | 1101 | AddressLiteral fd_toc(fd->toc()); |
goetz@6458 | 1102 | load_toc_from_toc(R2_TOC, fd_toc, toc); |
goetz@6458 | 1103 | // R2_TOC is killed. |
goetz@6458 | 1104 | bctrl(); |
goetz@6458 | 1105 | _last_calls_return_pc = pc(); |
goetz@6458 | 1106 | } else { |
goetz@6458 | 1107 | // It's a friend function, load the entry point and don't care about |
goetz@6458 | 1108 | // toc and env. Use an optimizable call instruction, but ensure the |
goetz@6458 | 1109 | // same code-size as in the case of a non-friend function. |
goetz@6458 | 1110 | nop(); |
goetz@6458 | 1111 | bl64_patchable(fd->entry(), rt); |
goetz@6458 | 1112 | _last_calls_return_pc = pc(); |
goetz@6458 | 1113 | } |
goetz@6458 | 1114 | return _last_calls_return_pc; |
goetz@6458 | 1115 | } |
goetz@6515 | 1116 | #endif // ABI_ELFv2 |
goetz@6458 | 1117 | |
goetz@6458 | 1118 | void MacroAssembler::call_VM_base(Register oop_result, |
goetz@6458 | 1119 | Register last_java_sp, |
goetz@6458 | 1120 | address entry_point, |
goetz@6458 | 1121 | bool check_exceptions) { |
goetz@6458 | 1122 | BLOCK_COMMENT("call_VM {"); |
goetz@6458 | 1123 | // Determine last_java_sp register. |
goetz@6458 | 1124 | if (!last_java_sp->is_valid()) { |
goetz@6458 | 1125 | last_java_sp = R1_SP; |
goetz@6458 | 1126 | } |
goetz@6458 | 1127 | set_top_ijava_frame_at_SP_as_last_Java_frame(last_java_sp, R11_scratch1); |
goetz@6458 | 1128 | |
goetz@6458 | 1129 | // ARG1 must hold thread address. |
goetz@6458 | 1130 | mr(R3_ARG1, R16_thread); |
goetz@6511 | 1131 | #if defined(ABI_ELFv2) |
goetz@6511 | 1132 | address return_pc = call_c(entry_point, relocInfo::none); |
goetz@6511 | 1133 | #else |
goetz@6458 | 1134 | address return_pc = call_c((FunctionDescriptor*)entry_point, relocInfo::none); |
goetz@6511 | 1135 | #endif |
goetz@6458 | 1136 | |
goetz@6458 | 1137 | reset_last_Java_frame(); |
goetz@6458 | 1138 | |
goetz@6458 | 1139 | // Check for pending exceptions. |
goetz@6458 | 1140 | if (check_exceptions) { |
goetz@6458 | 1141 | // We don't check for exceptions here. |
goetz@6458 | 1142 | ShouldNotReachHere(); |
goetz@6458 | 1143 | } |
goetz@6458 | 1144 | |
goetz@6458 | 1145 | // Get oop result if there is one and reset the value in the thread. |
goetz@6458 | 1146 | if (oop_result->is_valid()) { |
goetz@6458 | 1147 | get_vm_result(oop_result); |
goetz@6458 | 1148 | } |
goetz@6458 | 1149 | |
goetz@6458 | 1150 | _last_calls_return_pc = return_pc; |
goetz@6458 | 1151 | BLOCK_COMMENT("} call_VM"); |
goetz@6458 | 1152 | } |
goetz@6458 | 1153 | |
goetz@6458 | 1154 | void MacroAssembler::call_VM_leaf_base(address entry_point) { |
goetz@6458 | 1155 | BLOCK_COMMENT("call_VM_leaf {"); |
goetz@6511 | 1156 | #if defined(ABI_ELFv2) |
goetz@6511 | 1157 | call_c(entry_point, relocInfo::none); |
goetz@6511 | 1158 | #else |
goetz@6458 | 1159 | call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, entry_point), relocInfo::none); |
goetz@6511 | 1160 | #endif |
goetz@6458 | 1161 | BLOCK_COMMENT("} call_VM_leaf"); |
goetz@6458 | 1162 | } |
goetz@6458 | 1163 | |
goetz@6458 | 1164 | void MacroAssembler::call_VM(Register oop_result, address entry_point, bool check_exceptions) { |
goetz@6458 | 1165 | call_VM_base(oop_result, noreg, entry_point, check_exceptions); |
goetz@6458 | 1166 | } |
goetz@6458 | 1167 | |
goetz@6458 | 1168 | void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, |
goetz@6458 | 1169 | bool check_exceptions) { |
goetz@6458 | 1170 | // R3_ARG1 is reserved for the thread. |
goetz@6458 | 1171 | mr_if_needed(R4_ARG2, arg_1); |
goetz@6458 | 1172 | call_VM(oop_result, entry_point, check_exceptions); |
goetz@6458 | 1173 | } |
goetz@6458 | 1174 | |
goetz@6458 | 1175 | void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2, |
goetz@6458 | 1176 | bool check_exceptions) { |
goetz@6458 | 1177 | // R3_ARG1 is reserved for the thread |
goetz@6458 | 1178 | mr_if_needed(R4_ARG2, arg_1); |
goetz@6458 | 1179 | assert(arg_2 != R4_ARG2, "smashed argument"); |
goetz@6458 | 1180 | mr_if_needed(R5_ARG3, arg_2); |
goetz@6458 | 1181 | call_VM(oop_result, entry_point, check_exceptions); |
goetz@6458 | 1182 | } |
goetz@6458 | 1183 | |
goetz@7424 | 1184 | void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2, Register arg_3, |
goetz@7424 | 1185 | bool check_exceptions) { |
goetz@7424 | 1186 | // R3_ARG1 is reserved for the thread |
goetz@7424 | 1187 | mr_if_needed(R4_ARG2, arg_1); |
goetz@7424 | 1188 | assert(arg_2 != R4_ARG2, "smashed argument"); |
goetz@7424 | 1189 | mr_if_needed(R5_ARG3, arg_2); |
goetz@7424 | 1190 | mr_if_needed(R6_ARG4, arg_3); |
goetz@7424 | 1191 | call_VM(oop_result, entry_point, check_exceptions); |
goetz@7424 | 1192 | } |
goetz@7424 | 1193 | |
goetz@6458 | 1194 | void MacroAssembler::call_VM_leaf(address entry_point) { |
goetz@6458 | 1195 | call_VM_leaf_base(entry_point); |
goetz@6458 | 1196 | } |
goetz@6458 | 1197 | |
goetz@6458 | 1198 | void MacroAssembler::call_VM_leaf(address entry_point, Register arg_1) { |
goetz@6458 | 1199 | mr_if_needed(R3_ARG1, arg_1); |
goetz@6458 | 1200 | call_VM_leaf(entry_point); |
goetz@6458 | 1201 | } |
goetz@6458 | 1202 | |
goetz@6458 | 1203 | void MacroAssembler::call_VM_leaf(address entry_point, Register arg_1, Register arg_2) { |
goetz@6458 | 1204 | mr_if_needed(R3_ARG1, arg_1); |
goetz@6458 | 1205 | assert(arg_2 != R3_ARG1, "smashed argument"); |
goetz@6458 | 1206 | mr_if_needed(R4_ARG2, arg_2); |
goetz@6458 | 1207 | call_VM_leaf(entry_point); |
goetz@6458 | 1208 | } |
goetz@6458 | 1209 | |
goetz@6458 | 1210 | void MacroAssembler::call_VM_leaf(address entry_point, Register arg_1, Register arg_2, Register arg_3) { |
goetz@6458 | 1211 | mr_if_needed(R3_ARG1, arg_1); |
goetz@6458 | 1212 | assert(arg_2 != R3_ARG1, "smashed argument"); |
goetz@6458 | 1213 | mr_if_needed(R4_ARG2, arg_2); |
goetz@6458 | 1214 | assert(arg_3 != R3_ARG1 && arg_3 != R4_ARG2, "smashed argument"); |
goetz@6458 | 1215 | mr_if_needed(R5_ARG3, arg_3); |
goetz@6458 | 1216 | call_VM_leaf(entry_point); |
goetz@6458 | 1217 | } |
goetz@6458 | 1218 | |
goetz@6458 | 1219 | // Check whether instruction is a read access to the polling page |
goetz@6458 | 1220 | // which was emitted by load_from_polling_page(..). |
goetz@6458 | 1221 | bool MacroAssembler::is_load_from_polling_page(int instruction, void* ucontext, |
goetz@6458 | 1222 | address* polling_address_ptr) { |
goetz@6458 | 1223 | if (!is_ld(instruction)) |
goetz@6458 | 1224 | return false; // It's not a ld. Fail. |
goetz@6458 | 1225 | |
goetz@6458 | 1226 | int rt = inv_rt_field(instruction); |
goetz@6458 | 1227 | int ra = inv_ra_field(instruction); |
goetz@6458 | 1228 | int ds = inv_ds_field(instruction); |
goetz@6458 | 1229 | if (!(ds == 0 && ra != 0 && rt == 0)) { |
goetz@6458 | 1230 | return false; // It's not a ld(r0, X, ra). Fail. |
goetz@6458 | 1231 | } |
goetz@6458 | 1232 | |
goetz@6458 | 1233 | if (!ucontext) { |
goetz@6458 | 1234 | // Set polling address. |
goetz@6458 | 1235 | if (polling_address_ptr != NULL) { |
goetz@6458 | 1236 | *polling_address_ptr = NULL; |
goetz@6458 | 1237 | } |
goetz@6458 | 1238 | return true; // No ucontext given. Can't check value of ra. Assume true. |
goetz@6458 | 1239 | } |
goetz@6458 | 1240 | |
goetz@6458 | 1241 | #ifdef LINUX |
goetz@6458 | 1242 | // Ucontext given. Check that register ra contains the address of |
goetz@6458 | 1243 | // the safepoing polling page. |
goetz@6458 | 1244 | ucontext_t* uc = (ucontext_t*) ucontext; |
goetz@6458 | 1245 | // Set polling address. |
goetz@6458 | 1246 | address addr = (address)uc->uc_mcontext.regs->gpr[ra] + (ssize_t)ds; |
goetz@6458 | 1247 | if (polling_address_ptr != NULL) { |
goetz@6458 | 1248 | *polling_address_ptr = addr; |
goetz@6458 | 1249 | } |
goetz@6458 | 1250 | return os::is_poll_address(addr); |
goetz@6458 | 1251 | #else |
goetz@6458 | 1252 | // Not on Linux, ucontext must be NULL. |
goetz@6458 | 1253 | ShouldNotReachHere(); |
goetz@6458 | 1254 | return false; |
goetz@6458 | 1255 | #endif |
goetz@6458 | 1256 | } |
goetz@6458 | 1257 | |
goetz@6458 | 1258 | bool MacroAssembler::is_memory_serialization(int instruction, JavaThread* thread, void* ucontext) { |
goetz@6458 | 1259 | #ifdef LINUX |
goetz@6458 | 1260 | ucontext_t* uc = (ucontext_t*) ucontext; |
goetz@6458 | 1261 | |
goetz@6458 | 1262 | if (is_stwx(instruction) || is_stwux(instruction)) { |
goetz@6458 | 1263 | int ra = inv_ra_field(instruction); |
goetz@6458 | 1264 | int rb = inv_rb_field(instruction); |
goetz@6458 | 1265 | |
goetz@6458 | 1266 | // look up content of ra and rb in ucontext |
goetz@6458 | 1267 | address ra_val=(address)uc->uc_mcontext.regs->gpr[ra]; |
goetz@6458 | 1268 | long rb_val=(long)uc->uc_mcontext.regs->gpr[rb]; |
goetz@6458 | 1269 | return os::is_memory_serialize_page(thread, ra_val+rb_val); |
goetz@6458 | 1270 | } else if (is_stw(instruction) || is_stwu(instruction)) { |
goetz@6458 | 1271 | int ra = inv_ra_field(instruction); |
goetz@6458 | 1272 | int d1 = inv_d1_field(instruction); |
goetz@6458 | 1273 | |
goetz@6458 | 1274 | // look up content of ra in ucontext |
goetz@6458 | 1275 | address ra_val=(address)uc->uc_mcontext.regs->gpr[ra]; |
goetz@6458 | 1276 | return os::is_memory_serialize_page(thread, ra_val+d1); |
goetz@6458 | 1277 | } else { |
goetz@6458 | 1278 | return false; |
goetz@6458 | 1279 | } |
goetz@6458 | 1280 | #else |
goetz@6458 | 1281 | // workaround not needed on !LINUX :-) |
goetz@6458 | 1282 | ShouldNotCallThis(); |
goetz@6458 | 1283 | return false; |
goetz@6458 | 1284 | #endif |
goetz@6458 | 1285 | } |
goetz@6458 | 1286 | |
goetz@6458 | 1287 | void MacroAssembler::bang_stack_with_offset(int offset) { |
goetz@6458 | 1288 | // When increasing the stack, the old stack pointer will be written |
goetz@6458 | 1289 | // to the new top of stack according to the PPC64 abi. |
goetz@6458 | 1290 | // Therefore, stack banging is not necessary when increasing |
goetz@6458 | 1291 | // the stack by <= os::vm_page_size() bytes. |
goetz@6458 | 1292 | // When increasing the stack by a larger amount, this method is |
goetz@6458 | 1293 | // called repeatedly to bang the intermediate pages. |
goetz@6458 | 1294 | |
goetz@6458 | 1295 | // Stack grows down, caller passes positive offset. |
goetz@6458 | 1296 | assert(offset > 0, "must bang with positive offset"); |
goetz@6458 | 1297 | |
goetz@6458 | 1298 | long stdoffset = -offset; |
goetz@6458 | 1299 | |
goetz@6458 | 1300 | if (is_simm(stdoffset, 16)) { |
goetz@6458 | 1301 | // Signed 16 bit offset, a simple std is ok. |
goetz@6458 | 1302 | if (UseLoadInstructionsForStackBangingPPC64) { |
goetz@6458 | 1303 | ld(R0, (int)(signed short)stdoffset, R1_SP); |
goetz@6458 | 1304 | } else { |
goetz@6458 | 1305 | std(R0,(int)(signed short)stdoffset, R1_SP); |
goetz@6458 | 1306 | } |
goetz@6458 | 1307 | } else if (is_simm(stdoffset, 31)) { |
goetz@6458 | 1308 | const int hi = MacroAssembler::largeoffset_si16_si16_hi(stdoffset); |
goetz@6458 | 1309 | const int lo = MacroAssembler::largeoffset_si16_si16_lo(stdoffset); |
goetz@6458 | 1310 | |
goetz@6458 | 1311 | Register tmp = R11; |
goetz@6458 | 1312 | addis(tmp, R1_SP, hi); |
goetz@6458 | 1313 | if (UseLoadInstructionsForStackBangingPPC64) { |
goetz@6458 | 1314 | ld(R0, lo, tmp); |
goetz@6458 | 1315 | } else { |
goetz@6458 | 1316 | std(R0, lo, tmp); |
goetz@6458 | 1317 | } |
goetz@6458 | 1318 | } else { |
goetz@6458 | 1319 | ShouldNotReachHere(); |
goetz@6458 | 1320 | } |
goetz@6458 | 1321 | } |
goetz@6458 | 1322 | |
goetz@6458 | 1323 | // If instruction is a stack bang of the form |
goetz@6458 | 1324 | // std R0, x(Ry), (see bang_stack_with_offset()) |
goetz@6458 | 1325 | // stdu R1_SP, x(R1_SP), (see push_frame(), resize_frame()) |
goetz@6458 | 1326 | // or stdux R1_SP, Rx, R1_SP (see push_frame(), resize_frame()) |
goetz@6458 | 1327 | // return the banged address. Otherwise, return 0. |
goetz@6458 | 1328 | address MacroAssembler::get_stack_bang_address(int instruction, void *ucontext) { |
goetz@6458 | 1329 | #ifdef LINUX |
goetz@6458 | 1330 | ucontext_t* uc = (ucontext_t*) ucontext; |
goetz@6458 | 1331 | int rs = inv_rs_field(instruction); |
goetz@6458 | 1332 | int ra = inv_ra_field(instruction); |
goetz@6458 | 1333 | if ( (is_ld(instruction) && rs == 0 && UseLoadInstructionsForStackBangingPPC64) |
goetz@6458 | 1334 | || (is_std(instruction) && rs == 0 && !UseLoadInstructionsForStackBangingPPC64) |
goetz@6458 | 1335 | || (is_stdu(instruction) && rs == 1)) { |
goetz@6458 | 1336 | int ds = inv_ds_field(instruction); |
goetz@6458 | 1337 | // return banged address |
goetz@6458 | 1338 | return ds+(address)uc->uc_mcontext.regs->gpr[ra]; |
goetz@6458 | 1339 | } else if (is_stdux(instruction) && rs == 1) { |
goetz@6458 | 1340 | int rb = inv_rb_field(instruction); |
goetz@6458 | 1341 | address sp = (address)uc->uc_mcontext.regs->gpr[1]; |
goetz@6458 | 1342 | long rb_val = (long)uc->uc_mcontext.regs->gpr[rb]; |
goetz@6458 | 1343 | return ra != 1 || rb_val >= 0 ? NULL // not a stack bang |
goetz@6458 | 1344 | : sp + rb_val; // banged address |
goetz@6458 | 1345 | } |
goetz@6458 | 1346 | return NULL; // not a stack bang |
goetz@6458 | 1347 | #else |
goetz@6458 | 1348 | // workaround not needed on !LINUX :-) |
goetz@6458 | 1349 | ShouldNotCallThis(); |
goetz@6458 | 1350 | return NULL; |
goetz@6458 | 1351 | #endif |
goetz@6458 | 1352 | } |
goetz@6458 | 1353 | |
goetz@6458 | 1354 | // CmpxchgX sets condition register to cmpX(current, compare). |
goetz@6458 | 1355 | void MacroAssembler::cmpxchgw(ConditionRegister flag, Register dest_current_value, |
goetz@6458 | 1356 | Register compare_value, Register exchange_value, |
goetz@6458 | 1357 | Register addr_base, int semantics, bool cmpxchgx_hint, |
goetz@6458 | 1358 | Register int_flag_success, bool contention_hint) { |
goetz@6458 | 1359 | Label retry; |
goetz@6458 | 1360 | Label failed; |
goetz@6458 | 1361 | Label done; |
goetz@6458 | 1362 | |
goetz@6458 | 1363 | // Save one branch if result is returned via register and |
goetz@6458 | 1364 | // result register is different from the other ones. |
goetz@6458 | 1365 | bool use_result_reg = (int_flag_success != noreg); |
goetz@6458 | 1366 | bool preset_result_reg = (int_flag_success != dest_current_value && int_flag_success != compare_value && |
goetz@6458 | 1367 | int_flag_success != exchange_value && int_flag_success != addr_base); |
goetz@6458 | 1368 | |
goetz@6458 | 1369 | // release/fence semantics |
goetz@6458 | 1370 | if (semantics & MemBarRel) { |
goetz@6458 | 1371 | release(); |
goetz@6458 | 1372 | } |
goetz@6458 | 1373 | |
goetz@6458 | 1374 | if (use_result_reg && preset_result_reg) { |
goetz@6458 | 1375 | li(int_flag_success, 0); // preset (assume cas failed) |
goetz@6458 | 1376 | } |
goetz@6458 | 1377 | |
goetz@6458 | 1378 | // Add simple guard in order to reduce risk of starving under high contention (recommended by IBM). |
goetz@6458 | 1379 | if (contention_hint) { // Don't try to reserve if cmp fails. |
goetz@6458 | 1380 | lwz(dest_current_value, 0, addr_base); |
goetz@6458 | 1381 | cmpw(flag, dest_current_value, compare_value); |
goetz@6458 | 1382 | bne(flag, failed); |
goetz@6458 | 1383 | } |
goetz@6458 | 1384 | |
goetz@6458 | 1385 | // atomic emulation loop |
goetz@6458 | 1386 | bind(retry); |
goetz@6458 | 1387 | |
goetz@6458 | 1388 | lwarx(dest_current_value, addr_base, cmpxchgx_hint); |
goetz@6458 | 1389 | cmpw(flag, dest_current_value, compare_value); |
goetz@6458 | 1390 | if (UseStaticBranchPredictionInCompareAndSwapPPC64) { |
goetz@6458 | 1391 | bne_predict_not_taken(flag, failed); |
goetz@6458 | 1392 | } else { |
goetz@6458 | 1393 | bne( flag, failed); |
goetz@6458 | 1394 | } |
goetz@6458 | 1395 | // branch to done => (flag == ne), (dest_current_value != compare_value) |
goetz@6458 | 1396 | // fall through => (flag == eq), (dest_current_value == compare_value) |
goetz@6458 | 1397 | |
goetz@6458 | 1398 | stwcx_(exchange_value, addr_base); |
goetz@6458 | 1399 | if (UseStaticBranchPredictionInCompareAndSwapPPC64) { |
goetz@6458 | 1400 | bne_predict_not_taken(CCR0, retry); // StXcx_ sets CCR0. |
goetz@6458 | 1401 | } else { |
goetz@6458 | 1402 | bne( CCR0, retry); // StXcx_ sets CCR0. |
goetz@6458 | 1403 | } |
goetz@6458 | 1404 | // fall through => (flag == eq), (dest_current_value == compare_value), (swapped) |
goetz@6458 | 1405 | |
goetz@6458 | 1406 | // Result in register (must do this at the end because int_flag_success can be the |
goetz@6458 | 1407 | // same register as one above). |
goetz@6458 | 1408 | if (use_result_reg) { |
goetz@6458 | 1409 | li(int_flag_success, 1); |
goetz@6458 | 1410 | } |
goetz@6458 | 1411 | |
goetz@6458 | 1412 | if (semantics & MemBarFenceAfter) { |
goetz@6458 | 1413 | fence(); |
goetz@6458 | 1414 | } else if (semantics & MemBarAcq) { |
goetz@6458 | 1415 | isync(); |
goetz@6458 | 1416 | } |
goetz@6458 | 1417 | |
goetz@6458 | 1418 | if (use_result_reg && !preset_result_reg) { |
goetz@6458 | 1419 | b(done); |
goetz@6458 | 1420 | } |
goetz@6458 | 1421 | |
goetz@6458 | 1422 | bind(failed); |
goetz@6458 | 1423 | if (use_result_reg && !preset_result_reg) { |
goetz@6458 | 1424 | li(int_flag_success, 0); |
goetz@6458 | 1425 | } |
goetz@6458 | 1426 | |
goetz@6458 | 1427 | bind(done); |
goetz@6458 | 1428 | // (flag == ne) => (dest_current_value != compare_value), (!swapped) |
goetz@6458 | 1429 | // (flag == eq) => (dest_current_value == compare_value), ( swapped) |
goetz@6458 | 1430 | } |
goetz@6458 | 1431 | |
goetz@6458 | 1432 | // Preforms atomic compare exchange: |
goetz@6458 | 1433 | // if (compare_value == *addr_base) |
goetz@6458 | 1434 | // *addr_base = exchange_value |
goetz@6458 | 1435 | // int_flag_success = 1; |
goetz@6458 | 1436 | // else |
goetz@6458 | 1437 | // int_flag_success = 0; |
goetz@6458 | 1438 | // |
goetz@6458 | 1439 | // ConditionRegister flag = cmp(compare_value, *addr_base) |
goetz@6458 | 1440 | // Register dest_current_value = *addr_base |
goetz@6458 | 1441 | // Register compare_value Used to compare with value in memory |
goetz@6458 | 1442 | // Register exchange_value Written to memory if compare_value == *addr_base |
goetz@6458 | 1443 | // Register addr_base The memory location to compareXChange |
goetz@6458 | 1444 | // Register int_flag_success Set to 1 if exchange_value was written to *addr_base |
goetz@6458 | 1445 | // |
goetz@6458 | 1446 | // To avoid the costly compare exchange the value is tested beforehand. |
goetz@6458 | 1447 | // Several special cases exist to avoid that unnecessary information is generated. |
goetz@6458 | 1448 | // |
goetz@6458 | 1449 | void MacroAssembler::cmpxchgd(ConditionRegister flag, |
goetz@6458 | 1450 | Register dest_current_value, Register compare_value, Register exchange_value, |
goetz@6458 | 1451 | Register addr_base, int semantics, bool cmpxchgx_hint, |
goetz@6458 | 1452 | Register int_flag_success, Label* failed_ext, bool contention_hint) { |
goetz@6458 | 1453 | Label retry; |
goetz@6458 | 1454 | Label failed_int; |
goetz@6458 | 1455 | Label& failed = (failed_ext != NULL) ? *failed_ext : failed_int; |
goetz@6458 | 1456 | Label done; |
goetz@6458 | 1457 | |
goetz@6458 | 1458 | // Save one branch if result is returned via register and result register is different from the other ones. |
goetz@6458 | 1459 | bool use_result_reg = (int_flag_success!=noreg); |
goetz@6458 | 1460 | bool preset_result_reg = (int_flag_success!=dest_current_value && int_flag_success!=compare_value && |
goetz@6458 | 1461 | int_flag_success!=exchange_value && int_flag_success!=addr_base); |
goetz@6458 | 1462 | assert(int_flag_success == noreg || failed_ext == NULL, "cannot have both"); |
goetz@6458 | 1463 | |
goetz@6458 | 1464 | // release/fence semantics |
goetz@6458 | 1465 | if (semantics & MemBarRel) { |
goetz@6458 | 1466 | release(); |
goetz@6458 | 1467 | } |
goetz@6458 | 1468 | |
goetz@6458 | 1469 | if (use_result_reg && preset_result_reg) { |
goetz@6458 | 1470 | li(int_flag_success, 0); // preset (assume cas failed) |
goetz@6458 | 1471 | } |
goetz@6458 | 1472 | |
goetz@6458 | 1473 | // Add simple guard in order to reduce risk of starving under high contention (recommended by IBM). |
goetz@6458 | 1474 | if (contention_hint) { // Don't try to reserve if cmp fails. |
goetz@6458 | 1475 | ld(dest_current_value, 0, addr_base); |
goetz@6458 | 1476 | cmpd(flag, dest_current_value, compare_value); |
goetz@6458 | 1477 | bne(flag, failed); |
goetz@6458 | 1478 | } |
goetz@6458 | 1479 | |
goetz@6458 | 1480 | // atomic emulation loop |
goetz@6458 | 1481 | bind(retry); |
goetz@6458 | 1482 | |
goetz@6458 | 1483 | ldarx(dest_current_value, addr_base, cmpxchgx_hint); |
goetz@6458 | 1484 | cmpd(flag, dest_current_value, compare_value); |
goetz@6458 | 1485 | if (UseStaticBranchPredictionInCompareAndSwapPPC64) { |
goetz@6458 | 1486 | bne_predict_not_taken(flag, failed); |
goetz@6458 | 1487 | } else { |
goetz@6458 | 1488 | bne( flag, failed); |
goetz@6458 | 1489 | } |
goetz@6458 | 1490 | |
goetz@6458 | 1491 | stdcx_(exchange_value, addr_base); |
goetz@6458 | 1492 | if (UseStaticBranchPredictionInCompareAndSwapPPC64) { |
goetz@6458 | 1493 | bne_predict_not_taken(CCR0, retry); // stXcx_ sets CCR0 |
goetz@6458 | 1494 | } else { |
goetz@6458 | 1495 | bne( CCR0, retry); // stXcx_ sets CCR0 |
goetz@6458 | 1496 | } |
goetz@6458 | 1497 | |
goetz@6458 | 1498 | // result in register (must do this at the end because int_flag_success can be the same register as one above) |
goetz@6458 | 1499 | if (use_result_reg) { |
goetz@6458 | 1500 | li(int_flag_success, 1); |
goetz@6458 | 1501 | } |
goetz@6458 | 1502 | |
goetz@6458 | 1503 | // POWER6 doesn't need isync in CAS. |
goetz@6458 | 1504 | // Always emit isync to be on the safe side. |
goetz@6458 | 1505 | if (semantics & MemBarFenceAfter) { |
goetz@6458 | 1506 | fence(); |
goetz@6458 | 1507 | } else if (semantics & MemBarAcq) { |
goetz@6458 | 1508 | isync(); |
goetz@6458 | 1509 | } |
goetz@6458 | 1510 | |
goetz@6458 | 1511 | if (use_result_reg && !preset_result_reg) { |
goetz@6458 | 1512 | b(done); |
goetz@6458 | 1513 | } |
goetz@6458 | 1514 | |
goetz@6458 | 1515 | bind(failed_int); |
goetz@6458 | 1516 | if (use_result_reg && !preset_result_reg) { |
goetz@6458 | 1517 | li(int_flag_success, 0); |
goetz@6458 | 1518 | } |
goetz@6458 | 1519 | |
goetz@6458 | 1520 | bind(done); |
goetz@6458 | 1521 | // (flag == ne) => (dest_current_value != compare_value), (!swapped) |
goetz@6458 | 1522 | // (flag == eq) => (dest_current_value == compare_value), ( swapped) |
goetz@6458 | 1523 | } |
goetz@6458 | 1524 | |
goetz@6458 | 1525 | // Look up the method for a megamorphic invokeinterface call. |
goetz@6458 | 1526 | // The target method is determined by <intf_klass, itable_index>. |
goetz@6458 | 1527 | // The receiver klass is in recv_klass. |
goetz@6458 | 1528 | // On success, the result will be in method_result, and execution falls through. |
goetz@6458 | 1529 | // On failure, execution transfers to the given label. |
goetz@6458 | 1530 | void MacroAssembler::lookup_interface_method(Register recv_klass, |
goetz@6458 | 1531 | Register intf_klass, |
goetz@6458 | 1532 | RegisterOrConstant itable_index, |
goetz@6458 | 1533 | Register method_result, |
goetz@6458 | 1534 | Register scan_temp, |
mdoerr@9034 | 1535 | Register temp2, |
mdoerr@9034 | 1536 | Label& L_no_such_interface, |
mdoerr@9034 | 1537 | bool return_method) { |
goetz@6458 | 1538 | assert_different_registers(recv_klass, intf_klass, method_result, scan_temp); |
goetz@6458 | 1539 | |
goetz@6458 | 1540 | // Compute start of first itableOffsetEntry (which is at the end of the vtable). |
goetz@6458 | 1541 | int vtable_base = InstanceKlass::vtable_start_offset() * wordSize; |
goetz@6458 | 1542 | int itentry_off = itableMethodEntry::method_offset_in_bytes(); |
goetz@6458 | 1543 | int logMEsize = exact_log2(itableMethodEntry::size() * wordSize); |
goetz@6458 | 1544 | int scan_step = itableOffsetEntry::size() * wordSize; |
goetz@6458 | 1545 | int log_vte_size= exact_log2(vtableEntry::size() * wordSize); |
goetz@6458 | 1546 | |
goetz@6458 | 1547 | lwz(scan_temp, InstanceKlass::vtable_length_offset() * wordSize, recv_klass); |
goetz@6458 | 1548 | // %%% We should store the aligned, prescaled offset in the klassoop. |
goetz@6458 | 1549 | // Then the next several instructions would fold away. |
goetz@6458 | 1550 | |
goetz@6458 | 1551 | sldi(scan_temp, scan_temp, log_vte_size); |
goetz@6458 | 1552 | addi(scan_temp, scan_temp, vtable_base); |
goetz@6458 | 1553 | add(scan_temp, recv_klass, scan_temp); |
goetz@6458 | 1554 | |
goetz@6458 | 1555 | // Adjust recv_klass by scaled itable_index, so we can free itable_index. |
mdoerr@9034 | 1556 | if (return_method) { |
mdoerr@9034 | 1557 | if (itable_index.is_register()) { |
mdoerr@9034 | 1558 | Register itable_offset = itable_index.as_register(); |
mdoerr@9034 | 1559 | sldi(method_result, itable_offset, logMEsize); |
mdoerr@9034 | 1560 | if (itentry_off) { addi(method_result, method_result, itentry_off); } |
mdoerr@9034 | 1561 | add(method_result, method_result, recv_klass); |
mdoerr@9034 | 1562 | } else { |
mdoerr@9034 | 1563 | long itable_offset = (long)itable_index.as_constant(); |
mdoerr@9034 | 1564 | // static address, no relocation |
mdoerr@9034 | 1565 | load_const_optimized(temp2, (itable_offset << logMEsize) + itentry_off); // static address, no relocation |
mdoerr@9034 | 1566 | add(method_result, temp2, recv_klass); |
mdoerr@9034 | 1567 | } |
goetz@6458 | 1568 | } |
goetz@6458 | 1569 | |
goetz@6458 | 1570 | // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) { |
goetz@6458 | 1571 | // if (scan->interface() == intf) { |
goetz@6458 | 1572 | // result = (klass + scan->offset() + itable_index); |
goetz@6458 | 1573 | // } |
goetz@6458 | 1574 | // } |
goetz@6458 | 1575 | Label search, found_method; |
goetz@6458 | 1576 | |
goetz@6458 | 1577 | for (int peel = 1; peel >= 0; peel--) { |
goetz@6458 | 1578 | // %%%% Could load both offset and interface in one ldx, if they were |
goetz@6458 | 1579 | // in the opposite order. This would save a load. |
mdoerr@9034 | 1580 | ld(temp2, itableOffsetEntry::interface_offset_in_bytes(), scan_temp); |
goetz@6458 | 1581 | |
goetz@6458 | 1582 | // Check that this entry is non-null. A null entry means that |
goetz@6458 | 1583 | // the receiver class doesn't implement the interface, and wasn't the |
goetz@6458 | 1584 | // same as when the caller was compiled. |
mdoerr@9034 | 1585 | cmpd(CCR0, temp2, intf_klass); |
goetz@6458 | 1586 | |
goetz@6458 | 1587 | if (peel) { |
goetz@6458 | 1588 | beq(CCR0, found_method); |
goetz@6458 | 1589 | } else { |
goetz@6458 | 1590 | bne(CCR0, search); |
goetz@6458 | 1591 | // (invert the test to fall through to found_method...) |
goetz@6458 | 1592 | } |
goetz@6458 | 1593 | |
goetz@6458 | 1594 | if (!peel) break; |
goetz@6458 | 1595 | |
goetz@6458 | 1596 | bind(search); |
goetz@6458 | 1597 | |
mdoerr@9034 | 1598 | cmpdi(CCR0, temp2, 0); |
goetz@6458 | 1599 | beq(CCR0, L_no_such_interface); |
goetz@6458 | 1600 | addi(scan_temp, scan_temp, scan_step); |
goetz@6458 | 1601 | } |
goetz@6458 | 1602 | |
goetz@6458 | 1603 | bind(found_method); |
goetz@6458 | 1604 | |
goetz@6458 | 1605 | // Got a hit. |
mdoerr@9034 | 1606 | if (return_method) { |
mdoerr@9034 | 1607 | int ito_offset = itableOffsetEntry::offset_offset_in_bytes(); |
mdoerr@9034 | 1608 | lwz(scan_temp, ito_offset, scan_temp); |
mdoerr@9034 | 1609 | ldx(method_result, scan_temp, method_result); |
mdoerr@9034 | 1610 | } |
goetz@6458 | 1611 | } |
goetz@6458 | 1612 | |
goetz@6458 | 1613 | // virtual method calling |
goetz@6458 | 1614 | void MacroAssembler::lookup_virtual_method(Register recv_klass, |
goetz@6458 | 1615 | RegisterOrConstant vtable_index, |
goetz@6458 | 1616 | Register method_result) { |
goetz@6458 | 1617 | |
goetz@6458 | 1618 | assert_different_registers(recv_klass, method_result, vtable_index.register_or_noreg()); |
goetz@6458 | 1619 | |
goetz@6458 | 1620 | const int base = InstanceKlass::vtable_start_offset() * wordSize; |
goetz@6458 | 1621 | assert(vtableEntry::size() * wordSize == wordSize, "adjust the scaling in the code below"); |
goetz@6458 | 1622 | |
goetz@6458 | 1623 | if (vtable_index.is_register()) { |
goetz@6458 | 1624 | sldi(vtable_index.as_register(), vtable_index.as_register(), LogBytesPerWord); |
goetz@6458 | 1625 | add(recv_klass, vtable_index.as_register(), recv_klass); |
goetz@6458 | 1626 | } else { |
goetz@6458 | 1627 | addi(recv_klass, recv_klass, vtable_index.as_constant() << LogBytesPerWord); |
goetz@6458 | 1628 | } |
goetz@6458 | 1629 | ld(R19_method, base + vtableEntry::method_offset_in_bytes(), recv_klass); |
goetz@6458 | 1630 | } |
goetz@6458 | 1631 | |
goetz@6458 | 1632 | /////////////////////////////////////////// subtype checking //////////////////////////////////////////// |
goetz@6458 | 1633 | |
goetz@6458 | 1634 | void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass, |
goetz@6458 | 1635 | Register super_klass, |
goetz@6458 | 1636 | Register temp1_reg, |
goetz@6458 | 1637 | Register temp2_reg, |
goetz@6458 | 1638 | Label& L_success, |
goetz@6458 | 1639 | Label& L_failure) { |
goetz@6458 | 1640 | |
goetz@6458 | 1641 | const Register check_cache_offset = temp1_reg; |
goetz@6458 | 1642 | const Register cached_super = temp2_reg; |
goetz@6458 | 1643 | |
goetz@6458 | 1644 | assert_different_registers(sub_klass, super_klass, check_cache_offset, cached_super); |
goetz@6458 | 1645 | |
goetz@6458 | 1646 | int sco_offset = in_bytes(Klass::super_check_offset_offset()); |
goetz@6458 | 1647 | int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); |
goetz@6458 | 1648 | |
goetz@6458 | 1649 | // If the pointers are equal, we are done (e.g., String[] elements). |
goetz@6458 | 1650 | // This self-check enables sharing of secondary supertype arrays among |
goetz@6458 | 1651 | // non-primary types such as array-of-interface. Otherwise, each such |
goetz@6458 | 1652 | // type would need its own customized SSA. |
goetz@6458 | 1653 | // We move this check to the front of the fast path because many |
goetz@6458 | 1654 | // type checks are in fact trivially successful in this manner, |
goetz@6458 | 1655 | // so we get a nicely predicted branch right at the start of the check. |
goetz@6458 | 1656 | cmpd(CCR0, sub_klass, super_klass); |
goetz@6458 | 1657 | beq(CCR0, L_success); |
goetz@6458 | 1658 | |
goetz@6458 | 1659 | // Check the supertype display: |
goetz@6458 | 1660 | lwz(check_cache_offset, sco_offset, super_klass); |
goetz@6458 | 1661 | // The loaded value is the offset from KlassOopDesc. |
goetz@6458 | 1662 | |
goetz@6458 | 1663 | ldx(cached_super, check_cache_offset, sub_klass); |
goetz@6458 | 1664 | cmpd(CCR0, cached_super, super_klass); |
goetz@6458 | 1665 | beq(CCR0, L_success); |
goetz@6458 | 1666 | |
goetz@6458 | 1667 | // This check has worked decisively for primary supers. |
goetz@6458 | 1668 | // Secondary supers are sought in the super_cache ('super_cache_addr'). |
goetz@6458 | 1669 | // (Secondary supers are interfaces and very deeply nested subtypes.) |
goetz@6458 | 1670 | // This works in the same check above because of a tricky aliasing |
goetz@6458 | 1671 | // between the super_cache and the primary super display elements. |
goetz@6458 | 1672 | // (The 'super_check_addr' can address either, as the case requires.) |
goetz@6458 | 1673 | // Note that the cache is updated below if it does not help us find |
goetz@6458 | 1674 | // what we need immediately. |
goetz@6458 | 1675 | // So if it was a primary super, we can just fail immediately. |
goetz@6458 | 1676 | // Otherwise, it's the slow path for us (no success at this point). |
goetz@6458 | 1677 | |
goetz@6458 | 1678 | cmpwi(CCR0, check_cache_offset, sc_offset); |
goetz@6458 | 1679 | bne(CCR0, L_failure); |
goetz@6458 | 1680 | // bind(slow_path); // fallthru |
goetz@6458 | 1681 | } |
goetz@6458 | 1682 | |
goetz@6458 | 1683 | void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass, |
goetz@6458 | 1684 | Register super_klass, |
goetz@6458 | 1685 | Register temp1_reg, |
goetz@6458 | 1686 | Register temp2_reg, |
goetz@6458 | 1687 | Label* L_success, |
goetz@6458 | 1688 | Register result_reg) { |
goetz@6458 | 1689 | const Register array_ptr = temp1_reg; // current value from cache array |
goetz@6458 | 1690 | const Register temp = temp2_reg; |
goetz@6458 | 1691 | |
goetz@6458 | 1692 | assert_different_registers(sub_klass, super_klass, array_ptr, temp); |
goetz@6458 | 1693 | |
goetz@6458 | 1694 | int source_offset = in_bytes(Klass::secondary_supers_offset()); |
goetz@6458 | 1695 | int target_offset = in_bytes(Klass::secondary_super_cache_offset()); |
goetz@6458 | 1696 | |
goetz@6458 | 1697 | int length_offset = Array<Klass*>::length_offset_in_bytes(); |
goetz@6458 | 1698 | int base_offset = Array<Klass*>::base_offset_in_bytes(); |
goetz@6458 | 1699 | |
goetz@6458 | 1700 | Label hit, loop, failure, fallthru; |
goetz@6458 | 1701 | |
goetz@6458 | 1702 | ld(array_ptr, source_offset, sub_klass); |
goetz@6458 | 1703 | |
goetz@6458 | 1704 | //assert(4 == arrayOopDesc::length_length_in_bytes(), "precondition violated."); |
goetz@6458 | 1705 | lwz(temp, length_offset, array_ptr); |
goetz@6458 | 1706 | cmpwi(CCR0, temp, 0); |
goetz@6458 | 1707 | beq(CCR0, result_reg!=noreg ? failure : fallthru); // length 0 |
goetz@6458 | 1708 | |
goetz@6458 | 1709 | mtctr(temp); // load ctr |
goetz@6458 | 1710 | |
goetz@6458 | 1711 | bind(loop); |
goetz@6458 | 1712 | // Oops in table are NO MORE compressed. |
goetz@6458 | 1713 | ld(temp, base_offset, array_ptr); |
goetz@6458 | 1714 | cmpd(CCR0, temp, super_klass); |
goetz@6458 | 1715 | beq(CCR0, hit); |
goetz@6458 | 1716 | addi(array_ptr, array_ptr, BytesPerWord); |
goetz@6458 | 1717 | bdnz(loop); |
goetz@6458 | 1718 | |
goetz@6458 | 1719 | bind(failure); |
goetz@6458 | 1720 | if (result_reg!=noreg) li(result_reg, 1); // load non-zero result (indicates a miss) |
goetz@6458 | 1721 | b(fallthru); |
goetz@6458 | 1722 | |
goetz@6458 | 1723 | bind(hit); |
goetz@6458 | 1724 | std(super_klass, target_offset, sub_klass); // save result to cache |
goetz@6458 | 1725 | if (result_reg != noreg) li(result_reg, 0); // load zero result (indicates a hit) |
goetz@6458 | 1726 | if (L_success != NULL) b(*L_success); |
goetz@6458 | 1727 | |
goetz@6458 | 1728 | bind(fallthru); |
goetz@6458 | 1729 | } |
goetz@6458 | 1730 | |
goetz@6458 | 1731 | // Try fast path, then go to slow one if not successful |
goetz@6458 | 1732 | void MacroAssembler::check_klass_subtype(Register sub_klass, |
goetz@6458 | 1733 | Register super_klass, |
goetz@6458 | 1734 | Register temp1_reg, |
goetz@6458 | 1735 | Register temp2_reg, |
goetz@6458 | 1736 | Label& L_success) { |
goetz@6458 | 1737 | Label L_failure; |
goetz@6458 | 1738 | check_klass_subtype_fast_path(sub_klass, super_klass, temp1_reg, temp2_reg, L_success, L_failure); |
goetz@6458 | 1739 | check_klass_subtype_slow_path(sub_klass, super_klass, temp1_reg, temp2_reg, &L_success); |
goetz@6458 | 1740 | bind(L_failure); // Fallthru if not successful. |
goetz@6458 | 1741 | } |
goetz@6458 | 1742 | |
goetz@6458 | 1743 | void MacroAssembler::check_method_handle_type(Register mtype_reg, Register mh_reg, |
goetz@6458 | 1744 | Register temp_reg, |
goetz@6458 | 1745 | Label& wrong_method_type) { |
goetz@6458 | 1746 | assert_different_registers(mtype_reg, mh_reg, temp_reg); |
goetz@6458 | 1747 | // Compare method type against that of the receiver. |
goetz@6458 | 1748 | load_heap_oop_not_null(temp_reg, delayed_value(java_lang_invoke_MethodHandle::type_offset_in_bytes, temp_reg), mh_reg); |
goetz@6458 | 1749 | cmpd(CCR0, temp_reg, mtype_reg); |
goetz@6458 | 1750 | bne(CCR0, wrong_method_type); |
goetz@6458 | 1751 | } |
goetz@6458 | 1752 | |
goetz@6458 | 1753 | RegisterOrConstant MacroAssembler::argument_offset(RegisterOrConstant arg_slot, |
goetz@6458 | 1754 | Register temp_reg, |
goetz@6458 | 1755 | int extra_slot_offset) { |
goetz@6458 | 1756 | // cf. TemplateTable::prepare_invoke(), if (load_receiver). |
goetz@6458 | 1757 | int stackElementSize = Interpreter::stackElementSize; |
goetz@6458 | 1758 | int offset = extra_slot_offset * stackElementSize; |
goetz@6458 | 1759 | if (arg_slot.is_constant()) { |
goetz@6458 | 1760 | offset += arg_slot.as_constant() * stackElementSize; |
goetz@6458 | 1761 | return offset; |
goetz@6458 | 1762 | } else { |
goetz@6458 | 1763 | assert(temp_reg != noreg, "must specify"); |
goetz@6458 | 1764 | sldi(temp_reg, arg_slot.as_register(), exact_log2(stackElementSize)); |
goetz@6458 | 1765 | if (offset != 0) |
goetz@6458 | 1766 | addi(temp_reg, temp_reg, offset); |
goetz@6458 | 1767 | return temp_reg; |
goetz@6458 | 1768 | } |
goetz@6458 | 1769 | } |
goetz@6458 | 1770 | |
goetz@6458 | 1771 | void MacroAssembler::biased_locking_enter(ConditionRegister cr_reg, Register obj_reg, |
goetz@6458 | 1772 | Register mark_reg, Register temp_reg, |
goetz@6458 | 1773 | Register temp2_reg, Label& done, Label* slow_case) { |
goetz@6458 | 1774 | assert(UseBiasedLocking, "why call this otherwise?"); |
goetz@6458 | 1775 | |
goetz@6458 | 1776 | #ifdef ASSERT |
goetz@6458 | 1777 | assert_different_registers(obj_reg, mark_reg, temp_reg, temp2_reg); |
goetz@6458 | 1778 | #endif |
goetz@6458 | 1779 | |
goetz@6458 | 1780 | Label cas_label; |
goetz@6458 | 1781 | |
goetz@6458 | 1782 | // Branch to done if fast path fails and no slow_case provided. |
goetz@6458 | 1783 | Label *slow_case_int = (slow_case != NULL) ? slow_case : &done; |
goetz@6458 | 1784 | |
goetz@6458 | 1785 | // Biased locking |
goetz@6458 | 1786 | // See whether the lock is currently biased toward our thread and |
goetz@6458 | 1787 | // whether the epoch is still valid |
goetz@6458 | 1788 | // Note that the runtime guarantees sufficient alignment of JavaThread |
goetz@6458 | 1789 | // pointers to allow age to be placed into low bits |
goetz@6458 | 1790 | assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, |
goetz@6458 | 1791 | "biased locking makes assumptions about bit layout"); |
goetz@6458 | 1792 | |
goetz@6458 | 1793 | if (PrintBiasedLockingStatistics) { |
goetz@6458 | 1794 | load_const(temp_reg, (address) BiasedLocking::total_entry_count_addr(), temp2_reg); |
goetz@6458 | 1795 | lwz(temp2_reg, 0, temp_reg); |
goetz@6458 | 1796 | addi(temp2_reg, temp2_reg, 1); |
goetz@6458 | 1797 | stw(temp2_reg, 0, temp_reg); |
goetz@6458 | 1798 | } |
goetz@6458 | 1799 | |
goetz@6458 | 1800 | andi(temp_reg, mark_reg, markOopDesc::biased_lock_mask_in_place); |
goetz@6458 | 1801 | cmpwi(cr_reg, temp_reg, markOopDesc::biased_lock_pattern); |
goetz@6458 | 1802 | bne(cr_reg, cas_label); |
goetz@6458 | 1803 | |
goetz@6515 | 1804 | load_klass(temp_reg, obj_reg); |
goetz@6458 | 1805 | |
goetz@6458 | 1806 | load_const_optimized(temp2_reg, ~((int) markOopDesc::age_mask_in_place)); |
goetz@6458 | 1807 | ld(temp_reg, in_bytes(Klass::prototype_header_offset()), temp_reg); |
goetz@6458 | 1808 | orr(temp_reg, R16_thread, temp_reg); |
goetz@6458 | 1809 | xorr(temp_reg, mark_reg, temp_reg); |
goetz@6458 | 1810 | andr(temp_reg, temp_reg, temp2_reg); |
goetz@6458 | 1811 | cmpdi(cr_reg, temp_reg, 0); |
goetz@6458 | 1812 | if (PrintBiasedLockingStatistics) { |
goetz@6458 | 1813 | Label l; |
goetz@6458 | 1814 | bne(cr_reg, l); |
goetz@6458 | 1815 | load_const(mark_reg, (address) BiasedLocking::biased_lock_entry_count_addr()); |
goetz@6458 | 1816 | lwz(temp2_reg, 0, mark_reg); |
goetz@6458 | 1817 | addi(temp2_reg, temp2_reg, 1); |
goetz@6458 | 1818 | stw(temp2_reg, 0, mark_reg); |
goetz@6458 | 1819 | // restore mark_reg |
goetz@6458 | 1820 | ld(mark_reg, oopDesc::mark_offset_in_bytes(), obj_reg); |
goetz@6458 | 1821 | bind(l); |
goetz@6458 | 1822 | } |
goetz@6458 | 1823 | beq(cr_reg, done); |
goetz@6458 | 1824 | |
goetz@6458 | 1825 | Label try_revoke_bias; |
goetz@6458 | 1826 | Label try_rebias; |
goetz@6458 | 1827 | |
goetz@6458 | 1828 | // At this point we know that the header has the bias pattern and |
goetz@6458 | 1829 | // that we are not the bias owner in the current epoch. We need to |
goetz@6458 | 1830 | // figure out more details about the state of the header in order to |
goetz@6458 | 1831 | // know what operations can be legally performed on the object's |
goetz@6458 | 1832 | // header. |
goetz@6458 | 1833 | |
goetz@6458 | 1834 | // If the low three bits in the xor result aren't clear, that means |
goetz@6458 | 1835 | // the prototype header is no longer biased and we have to revoke |
goetz@6458 | 1836 | // the bias on this object. |
goetz@6458 | 1837 | andi(temp2_reg, temp_reg, markOopDesc::biased_lock_mask_in_place); |
goetz@6458 | 1838 | cmpwi(cr_reg, temp2_reg, 0); |
goetz@6458 | 1839 | bne(cr_reg, try_revoke_bias); |
goetz@6458 | 1840 | |
goetz@6458 | 1841 | // Biasing is still enabled for this data type. See whether the |
goetz@6458 | 1842 | // epoch of the current bias is still valid, meaning that the epoch |
goetz@6458 | 1843 | // bits of the mark word are equal to the epoch bits of the |
goetz@6458 | 1844 | // prototype header. (Note that the prototype header's epoch bits |
goetz@6458 | 1845 | // only change at a safepoint.) If not, attempt to rebias the object |
goetz@6458 | 1846 | // toward the current thread. Note that we must be absolutely sure |
goetz@6458 | 1847 | // that the current epoch is invalid in order to do this because |
goetz@6458 | 1848 | // otherwise the manipulations it performs on the mark word are |
goetz@6458 | 1849 | // illegal. |
goetz@6458 | 1850 | |
goetz@6458 | 1851 | int shift_amount = 64 - markOopDesc::epoch_shift; |
goetz@6458 | 1852 | // rotate epoch bits to right (little) end and set other bits to 0 |
goetz@6458 | 1853 | // [ big part | epoch | little part ] -> [ 0..0 | epoch ] |
goetz@6458 | 1854 | rldicl_(temp2_reg, temp_reg, shift_amount, 64 - markOopDesc::epoch_bits); |
goetz@6458 | 1855 | // branch if epoch bits are != 0, i.e. they differ, because the epoch has been incremented |
goetz@6458 | 1856 | bne(CCR0, try_rebias); |
goetz@6458 | 1857 | |
goetz@6458 | 1858 | // The epoch of the current bias is still valid but we know nothing |
goetz@6458 | 1859 | // about the owner; it might be set or it might be clear. Try to |
goetz@6458 | 1860 | // acquire the bias of the object using an atomic operation. If this |
goetz@6458 | 1861 | // fails we will go in to the runtime to revoke the object's bias. |
goetz@6458 | 1862 | // Note that we first construct the presumed unbiased header so we |
goetz@6458 | 1863 | // don't accidentally blow away another thread's valid bias. |
goetz@6458 | 1864 | andi(mark_reg, mark_reg, (markOopDesc::biased_lock_mask_in_place | |
goetz@6458 | 1865 | markOopDesc::age_mask_in_place | |
goetz@6458 | 1866 | markOopDesc::epoch_mask_in_place)); |
goetz@6458 | 1867 | orr(temp_reg, R16_thread, mark_reg); |
goetz@6458 | 1868 | |
goetz@6458 | 1869 | assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0"); |
goetz@6458 | 1870 | |
goetz@6458 | 1871 | // CmpxchgX sets cr_reg to cmpX(temp2_reg, mark_reg). |
goetz@6458 | 1872 | fence(); // TODO: replace by MacroAssembler::MemBarRel | MacroAssembler::MemBarAcq ? |
goetz@6458 | 1873 | cmpxchgd(/*flag=*/cr_reg, /*current_value=*/temp2_reg, |
goetz@6458 | 1874 | /*compare_value=*/mark_reg, /*exchange_value=*/temp_reg, |
goetz@6458 | 1875 | /*where=*/obj_reg, |
goetz@6458 | 1876 | MacroAssembler::MemBarAcq, |
goetz@6458 | 1877 | MacroAssembler::cmpxchgx_hint_acquire_lock(), |
goetz@6458 | 1878 | noreg, slow_case_int); // bail out if failed |
goetz@6458 | 1879 | |
goetz@6458 | 1880 | // If the biasing toward our thread failed, this means that |
goetz@6458 | 1881 | // another thread succeeded in biasing it toward itself and we |
goetz@6458 | 1882 | // need to revoke that bias. The revocation will occur in the |
goetz@6458 | 1883 | // interpreter runtime in the slow case. |
goetz@6458 | 1884 | if (PrintBiasedLockingStatistics) { |
goetz@6458 | 1885 | load_const(temp_reg, (address) BiasedLocking::anonymously_biased_lock_entry_count_addr(), temp2_reg); |
goetz@6458 | 1886 | lwz(temp2_reg, 0, temp_reg); |
goetz@6458 | 1887 | addi(temp2_reg, temp2_reg, 1); |
goetz@6458 | 1888 | stw(temp2_reg, 0, temp_reg); |
goetz@6458 | 1889 | } |
goetz@6458 | 1890 | b(done); |
goetz@6458 | 1891 | |
goetz@6458 | 1892 | bind(try_rebias); |
goetz@6458 | 1893 | // At this point we know the epoch has expired, meaning that the |
goetz@6458 | 1894 | // current "bias owner", if any, is actually invalid. Under these |
goetz@6458 | 1895 | // circumstances _only_, we are allowed to use the current header's |
goetz@6458 | 1896 | // value as the comparison value when doing the cas to acquire the |
goetz@6458 | 1897 | // bias in the current epoch. In other words, we allow transfer of |
goetz@6458 | 1898 | // the bias from one thread to another directly in this situation. |
goetz@6458 | 1899 | andi(temp_reg, mark_reg, markOopDesc::age_mask_in_place); |
goetz@6458 | 1900 | orr(temp_reg, R16_thread, temp_reg); |
goetz@6515 | 1901 | load_klass(temp2_reg, obj_reg); |
goetz@6458 | 1902 | ld(temp2_reg, in_bytes(Klass::prototype_header_offset()), temp2_reg); |
goetz@6458 | 1903 | orr(temp_reg, temp_reg, temp2_reg); |
goetz@6458 | 1904 | |
goetz@6458 | 1905 | assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0"); |
goetz@6458 | 1906 | |
goetz@6458 | 1907 | // CmpxchgX sets cr_reg to cmpX(temp2_reg, mark_reg). |
goetz@6458 | 1908 | fence(); // TODO: replace by MacroAssembler::MemBarRel | MacroAssembler::MemBarAcq ? |
goetz@6458 | 1909 | cmpxchgd(/*flag=*/cr_reg, /*current_value=*/temp2_reg, |
goetz@6458 | 1910 | /*compare_value=*/mark_reg, /*exchange_value=*/temp_reg, |
goetz@6458 | 1911 | /*where=*/obj_reg, |
goetz@6458 | 1912 | MacroAssembler::MemBarAcq, |
goetz@6458 | 1913 | MacroAssembler::cmpxchgx_hint_acquire_lock(), |
goetz@6458 | 1914 | noreg, slow_case_int); // bail out if failed |
goetz@6458 | 1915 | |
goetz@6458 | 1916 | // If the biasing toward our thread failed, this means that |
goetz@6458 | 1917 | // another thread succeeded in biasing it toward itself and we |
goetz@6458 | 1918 | // need to revoke that bias. The revocation will occur in the |
goetz@6458 | 1919 | // interpreter runtime in the slow case. |
goetz@6458 | 1920 | if (PrintBiasedLockingStatistics) { |
goetz@6458 | 1921 | load_const(temp_reg, (address) BiasedLocking::rebiased_lock_entry_count_addr(), temp2_reg); |
goetz@6458 | 1922 | lwz(temp2_reg, 0, temp_reg); |
goetz@6458 | 1923 | addi(temp2_reg, temp2_reg, 1); |
goetz@6458 | 1924 | stw(temp2_reg, 0, temp_reg); |
goetz@6458 | 1925 | } |
goetz@6458 | 1926 | b(done); |
goetz@6458 | 1927 | |
goetz@6458 | 1928 | bind(try_revoke_bias); |
goetz@6458 | 1929 | // The prototype mark in the klass doesn't have the bias bit set any |
goetz@6458 | 1930 | // more, indicating that objects of this data type are not supposed |
goetz@6458 | 1931 | // to be biased any more. We are going to try to reset the mark of |
goetz@6458 | 1932 | // this object to the prototype value and fall through to the |
goetz@6458 | 1933 | // CAS-based locking scheme. Note that if our CAS fails, it means |
goetz@6458 | 1934 | // that another thread raced us for the privilege of revoking the |
goetz@6458 | 1935 | // bias of this particular object, so it's okay to continue in the |
goetz@6458 | 1936 | // normal locking code. |
goetz@6515 | 1937 | load_klass(temp_reg, obj_reg); |
goetz@6458 | 1938 | ld(temp_reg, in_bytes(Klass::prototype_header_offset()), temp_reg); |
goetz@6458 | 1939 | andi(temp2_reg, mark_reg, markOopDesc::age_mask_in_place); |
goetz@6458 | 1940 | orr(temp_reg, temp_reg, temp2_reg); |
goetz@6458 | 1941 | |
goetz@6458 | 1942 | assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0"); |
goetz@6458 | 1943 | |
goetz@6458 | 1944 | // CmpxchgX sets cr_reg to cmpX(temp2_reg, mark_reg). |
goetz@6458 | 1945 | fence(); // TODO: replace by MacroAssembler::MemBarRel | MacroAssembler::MemBarAcq ? |
goetz@6458 | 1946 | cmpxchgd(/*flag=*/cr_reg, /*current_value=*/temp2_reg, |
goetz@6458 | 1947 | /*compare_value=*/mark_reg, /*exchange_value=*/temp_reg, |
goetz@6458 | 1948 | /*where=*/obj_reg, |
goetz@6458 | 1949 | MacroAssembler::MemBarAcq, |
goetz@6458 | 1950 | MacroAssembler::cmpxchgx_hint_acquire_lock()); |
goetz@6458 | 1951 | |
goetz@6458 | 1952 | // reload markOop in mark_reg before continuing with lightweight locking |
goetz@6458 | 1953 | ld(mark_reg, oopDesc::mark_offset_in_bytes(), obj_reg); |
goetz@6458 | 1954 | |
goetz@6458 | 1955 | // Fall through to the normal CAS-based lock, because no matter what |
goetz@6458 | 1956 | // the result of the above CAS, some thread must have succeeded in |
goetz@6458 | 1957 | // removing the bias bit from the object's header. |
goetz@6458 | 1958 | if (PrintBiasedLockingStatistics) { |
goetz@6458 | 1959 | Label l; |
goetz@6458 | 1960 | bne(cr_reg, l); |
goetz@6458 | 1961 | load_const(temp_reg, (address) BiasedLocking::revoked_lock_entry_count_addr(), temp2_reg); |
goetz@6458 | 1962 | lwz(temp2_reg, 0, temp_reg); |
goetz@6458 | 1963 | addi(temp2_reg, temp2_reg, 1); |
goetz@6458 | 1964 | stw(temp2_reg, 0, temp_reg); |
goetz@6458 | 1965 | bind(l); |
goetz@6458 | 1966 | } |
goetz@6458 | 1967 | |
goetz@6458 | 1968 | bind(cas_label); |
goetz@6458 | 1969 | } |
goetz@6458 | 1970 | |
goetz@6458 | 1971 | void MacroAssembler::biased_locking_exit (ConditionRegister cr_reg, Register mark_addr, Register temp_reg, Label& done) { |
goetz@6458 | 1972 | // Check for biased locking unlock case, which is a no-op |
goetz@6458 | 1973 | // Note: we do not have to check the thread ID for two reasons. |
goetz@6458 | 1974 | // First, the interpreter checks for IllegalMonitorStateException at |
goetz@6458 | 1975 | // a higher level. Second, if the bias was revoked while we held the |
goetz@6458 | 1976 | // lock, the object could not be rebiased toward another thread, so |
goetz@6458 | 1977 | // the bias bit would be clear. |
goetz@6458 | 1978 | |
goetz@6458 | 1979 | ld(temp_reg, 0, mark_addr); |
goetz@6458 | 1980 | andi(temp_reg, temp_reg, markOopDesc::biased_lock_mask_in_place); |
goetz@6458 | 1981 | |
goetz@6458 | 1982 | cmpwi(cr_reg, temp_reg, markOopDesc::biased_lock_pattern); |
goetz@6458 | 1983 | beq(cr_reg, done); |
goetz@6458 | 1984 | } |
goetz@6458 | 1985 | |
goetz@6458 | 1986 | // "The box" is the space on the stack where we copy the object mark. |
goetz@6458 | 1987 | void MacroAssembler::compiler_fast_lock_object(ConditionRegister flag, Register oop, Register box, |
goetz@6458 | 1988 | Register temp, Register displaced_header, Register current_header) { |
goetz@6458 | 1989 | assert_different_registers(oop, box, temp, displaced_header, current_header); |
goetz@6458 | 1990 | assert(flag != CCR0, "bad condition register"); |
goetz@6458 | 1991 | Label cont; |
goetz@6458 | 1992 | Label object_has_monitor; |
goetz@6458 | 1993 | Label cas_failed; |
goetz@6458 | 1994 | |
goetz@6458 | 1995 | // Load markOop from object into displaced_header. |
goetz@6458 | 1996 | ld(displaced_header, oopDesc::mark_offset_in_bytes(), oop); |
goetz@6458 | 1997 | |
goetz@6458 | 1998 | |
goetz@6458 | 1999 | // Always do locking in runtime. |
goetz@6458 | 2000 | if (EmitSync & 0x01) { |
goetz@6458 | 2001 | cmpdi(flag, oop, 0); // Oop can't be 0 here => always false. |
goetz@6458 | 2002 | return; |
goetz@6458 | 2003 | } |
goetz@6458 | 2004 | |
goetz@6458 | 2005 | if (UseBiasedLocking) { |
goetz@6458 | 2006 | biased_locking_enter(flag, oop, displaced_header, temp, current_header, cont); |
goetz@6458 | 2007 | } |
goetz@6458 | 2008 | |
goetz@6458 | 2009 | // Handle existing monitor. |
goetz@6458 | 2010 | if ((EmitSync & 0x02) == 0) { |
goetz@6458 | 2011 | // The object has an existing monitor iff (mark & monitor_value) != 0. |
goetz@6458 | 2012 | andi_(temp, displaced_header, markOopDesc::monitor_value); |
goetz@6458 | 2013 | bne(CCR0, object_has_monitor); |
goetz@6458 | 2014 | } |
goetz@6458 | 2015 | |
goetz@6458 | 2016 | // Set displaced_header to be (markOop of object | UNLOCK_VALUE). |
goetz@6458 | 2017 | ori(displaced_header, displaced_header, markOopDesc::unlocked_value); |
goetz@6458 | 2018 | |
goetz@6458 | 2019 | // Load Compare Value application register. |
goetz@6458 | 2020 | |
goetz@6458 | 2021 | // Initialize the box. (Must happen before we update the object mark!) |
goetz@6458 | 2022 | std(displaced_header, BasicLock::displaced_header_offset_in_bytes(), box); |
goetz@6458 | 2023 | |
goetz@6458 | 2024 | // Must fence, otherwise, preceding store(s) may float below cmpxchg. |
goetz@6458 | 2025 | // Compare object markOop with mark and if equal exchange scratch1 with object markOop. |
goetz@6458 | 2026 | // CmpxchgX sets cr_reg to cmpX(current, displaced). |
goetz@6501 | 2027 | membar(Assembler::StoreStore); |
goetz@6458 | 2028 | cmpxchgd(/*flag=*/flag, |
goetz@6458 | 2029 | /*current_value=*/current_header, |
goetz@6458 | 2030 | /*compare_value=*/displaced_header, |
goetz@6458 | 2031 | /*exchange_value=*/box, |
goetz@6458 | 2032 | /*where=*/oop, |
goetz@6501 | 2033 | MacroAssembler::MemBarAcq, |
goetz@6458 | 2034 | MacroAssembler::cmpxchgx_hint_acquire_lock(), |
goetz@6458 | 2035 | noreg, |
goetz@6458 | 2036 | &cas_failed); |
goetz@6458 | 2037 | assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0"); |
goetz@6458 | 2038 | |
goetz@6458 | 2039 | // If the compare-and-exchange succeeded, then we found an unlocked |
goetz@6458 | 2040 | // object and we have now locked it. |
goetz@6458 | 2041 | b(cont); |
goetz@6458 | 2042 | |
goetz@6458 | 2043 | bind(cas_failed); |
goetz@6458 | 2044 | // We did not see an unlocked object so try the fast recursive case. |
goetz@6458 | 2045 | |
goetz@6458 | 2046 | // Check if the owner is self by comparing the value in the markOop of object |
goetz@6458 | 2047 | // (current_header) with the stack pointer. |
goetz@6458 | 2048 | sub(current_header, current_header, R1_SP); |
goetz@6458 | 2049 | load_const_optimized(temp, (address) (~(os::vm_page_size()-1) | |
goetz@6458 | 2050 | markOopDesc::lock_mask_in_place)); |
goetz@6458 | 2051 | |
goetz@6458 | 2052 | and_(R0/*==0?*/, current_header, temp); |
goetz@6458 | 2053 | // If condition is true we are cont and hence we can store 0 as the |
goetz@6458 | 2054 | // displaced header in the box, which indicates that it is a recursive lock. |
goetz@6458 | 2055 | mcrf(flag,CCR0); |
goetz@6458 | 2056 | std(R0/*==0, perhaps*/, BasicLock::displaced_header_offset_in_bytes(), box); |
goetz@6458 | 2057 | |
goetz@6458 | 2058 | // Handle existing monitor. |
goetz@6458 | 2059 | if ((EmitSync & 0x02) == 0) { |
goetz@6458 | 2060 | b(cont); |
goetz@6458 | 2061 | |
goetz@6458 | 2062 | bind(object_has_monitor); |
goetz@6458 | 2063 | // The object's monitor m is unlocked iff m->owner == NULL, |
goetz@6458 | 2064 | // otherwise m->owner may contain a thread or a stack address. |
goetz@6458 | 2065 | // |
goetz@6458 | 2066 | // Try to CAS m->owner from NULL to current thread. |
goetz@6458 | 2067 | addi(temp, displaced_header, ObjectMonitor::owner_offset_in_bytes()-markOopDesc::monitor_value); |
goetz@6458 | 2068 | li(displaced_header, 0); |
goetz@6458 | 2069 | // CmpxchgX sets flag to cmpX(current, displaced). |
goetz@6458 | 2070 | cmpxchgd(/*flag=*/flag, |
goetz@6458 | 2071 | /*current_value=*/current_header, |
goetz@6458 | 2072 | /*compare_value=*/displaced_header, |
goetz@6458 | 2073 | /*exchange_value=*/R16_thread, |
goetz@6458 | 2074 | /*where=*/temp, |
goetz@6458 | 2075 | MacroAssembler::MemBarRel | MacroAssembler::MemBarAcq, |
goetz@6458 | 2076 | MacroAssembler::cmpxchgx_hint_acquire_lock()); |
goetz@6458 | 2077 | |
goetz@6458 | 2078 | // Store a non-null value into the box. |
goetz@6458 | 2079 | std(box, BasicLock::displaced_header_offset_in_bytes(), box); |
goetz@6458 | 2080 | |
goetz@6458 | 2081 | # ifdef ASSERT |
goetz@6458 | 2082 | bne(flag, cont); |
goetz@6458 | 2083 | // We have acquired the monitor, check some invariants. |
goetz@6458 | 2084 | addi(/*monitor=*/temp, temp, -ObjectMonitor::owner_offset_in_bytes()); |
goetz@6458 | 2085 | // Invariant 1: _recursions should be 0. |
goetz@6458 | 2086 | //assert(ObjectMonitor::recursions_size_in_bytes() == 8, "unexpected size"); |
goetz@6458 | 2087 | asm_assert_mem8_is_zero(ObjectMonitor::recursions_offset_in_bytes(), temp, |
goetz@6458 | 2088 | "monitor->_recursions should be 0", -1); |
goetz@6458 | 2089 | // Invariant 2: OwnerIsThread shouldn't be 0. |
goetz@6458 | 2090 | //assert(ObjectMonitor::OwnerIsThread_size_in_bytes() == 4, "unexpected size"); |
goetz@6458 | 2091 | //asm_assert_mem4_isnot_zero(ObjectMonitor::OwnerIsThread_offset_in_bytes(), temp, |
goetz@6458 | 2092 | // "monitor->OwnerIsThread shouldn't be 0", -1); |
goetz@6458 | 2093 | # endif |
goetz@6458 | 2094 | } |
goetz@6458 | 2095 | |
goetz@6458 | 2096 | bind(cont); |
goetz@6458 | 2097 | // flag == EQ indicates success |
goetz@6458 | 2098 | // flag == NE indicates failure |
goetz@6458 | 2099 | } |
goetz@6458 | 2100 | |
goetz@6458 | 2101 | void MacroAssembler::compiler_fast_unlock_object(ConditionRegister flag, Register oop, Register box, |
goetz@6458 | 2102 | Register temp, Register displaced_header, Register current_header) { |
goetz@6458 | 2103 | assert_different_registers(oop, box, temp, displaced_header, current_header); |
goetz@6458 | 2104 | assert(flag != CCR0, "bad condition register"); |
goetz@6458 | 2105 | Label cont; |
goetz@6458 | 2106 | Label object_has_monitor; |
goetz@6458 | 2107 | |
goetz@6458 | 2108 | // Always do locking in runtime. |
goetz@6458 | 2109 | if (EmitSync & 0x01) { |
goetz@6458 | 2110 | cmpdi(flag, oop, 0); // Oop can't be 0 here => always false. |
goetz@6458 | 2111 | return; |
goetz@6458 | 2112 | } |
goetz@6458 | 2113 | |
goetz@6458 | 2114 | if (UseBiasedLocking) { |
goetz@6458 | 2115 | biased_locking_exit(flag, oop, current_header, cont); |
goetz@6458 | 2116 | } |
goetz@6458 | 2117 | |
goetz@6458 | 2118 | // Find the lock address and load the displaced header from the stack. |
goetz@6458 | 2119 | ld(displaced_header, BasicLock::displaced_header_offset_in_bytes(), box); |
goetz@6458 | 2120 | |
goetz@6458 | 2121 | // If the displaced header is 0, we have a recursive unlock. |
goetz@6458 | 2122 | cmpdi(flag, displaced_header, 0); |
goetz@6458 | 2123 | beq(flag, cont); |
goetz@6458 | 2124 | |
goetz@6458 | 2125 | // Handle existing monitor. |
goetz@6458 | 2126 | if ((EmitSync & 0x02) == 0) { |
goetz@6458 | 2127 | // The object has an existing monitor iff (mark & monitor_value) != 0. |
goetz@6458 | 2128 | ld(current_header, oopDesc::mark_offset_in_bytes(), oop); |
goetz@6458 | 2129 | andi(temp, current_header, markOopDesc::monitor_value); |
goetz@6458 | 2130 | cmpdi(flag, temp, 0); |
goetz@6458 | 2131 | bne(flag, object_has_monitor); |
goetz@6458 | 2132 | } |
goetz@6458 | 2133 | |
goetz@6458 | 2134 | |
goetz@6458 | 2135 | // Check if it is still a light weight lock, this is is true if we see |
goetz@6458 | 2136 | // the stack address of the basicLock in the markOop of the object. |
goetz@6458 | 2137 | // Cmpxchg sets flag to cmpd(current_header, box). |
goetz@6458 | 2138 | cmpxchgd(/*flag=*/flag, |
goetz@6458 | 2139 | /*current_value=*/current_header, |
goetz@6458 | 2140 | /*compare_value=*/box, |
goetz@6458 | 2141 | /*exchange_value=*/displaced_header, |
goetz@6458 | 2142 | /*where=*/oop, |
goetz@6458 | 2143 | MacroAssembler::MemBarRel, |
goetz@6458 | 2144 | MacroAssembler::cmpxchgx_hint_release_lock(), |
goetz@6458 | 2145 | noreg, |
goetz@6458 | 2146 | &cont); |
goetz@6458 | 2147 | |
goetz@6458 | 2148 | assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0"); |
goetz@6458 | 2149 | |
goetz@6458 | 2150 | // Handle existing monitor. |
goetz@6458 | 2151 | if ((EmitSync & 0x02) == 0) { |
goetz@6458 | 2152 | b(cont); |
goetz@6458 | 2153 | |
goetz@6458 | 2154 | bind(object_has_monitor); |
goetz@6458 | 2155 | addi(current_header, current_header, -markOopDesc::monitor_value); // monitor |
goetz@6458 | 2156 | ld(temp, ObjectMonitor::owner_offset_in_bytes(), current_header); |
goetz@6458 | 2157 | ld(displaced_header, ObjectMonitor::recursions_offset_in_bytes(), current_header); |
goetz@6458 | 2158 | xorr(temp, R16_thread, temp); // Will be 0 if we are the owner. |
goetz@6458 | 2159 | orr(temp, temp, displaced_header); // Will be 0 if there are 0 recursions. |
goetz@6458 | 2160 | cmpdi(flag, temp, 0); |
goetz@6458 | 2161 | bne(flag, cont); |
goetz@6458 | 2162 | |
goetz@6458 | 2163 | ld(temp, ObjectMonitor::EntryList_offset_in_bytes(), current_header); |
goetz@6458 | 2164 | ld(displaced_header, ObjectMonitor::cxq_offset_in_bytes(), current_header); |
goetz@6458 | 2165 | orr(temp, temp, displaced_header); // Will be 0 if both are 0. |
goetz@6458 | 2166 | cmpdi(flag, temp, 0); |
goetz@6458 | 2167 | bne(flag, cont); |
goetz@6458 | 2168 | release(); |
goetz@6458 | 2169 | std(temp, ObjectMonitor::owner_offset_in_bytes(), current_header); |
goetz@6458 | 2170 | } |
goetz@6458 | 2171 | |
goetz@6458 | 2172 | bind(cont); |
goetz@6458 | 2173 | // flag == EQ indicates success |
goetz@6458 | 2174 | // flag == NE indicates failure |
goetz@6458 | 2175 | } |
goetz@6458 | 2176 | |
goetz@6458 | 2177 | // Write serialization page so VM thread can do a pseudo remote membar. |
goetz@6458 | 2178 | // We use the current thread pointer to calculate a thread specific |
goetz@6458 | 2179 | // offset to write to within the page. This minimizes bus traffic |
goetz@6458 | 2180 | // due to cache line collision. |
goetz@6458 | 2181 | void MacroAssembler::serialize_memory(Register thread, Register tmp1, Register tmp2) { |
goetz@6458 | 2182 | srdi(tmp2, thread, os::get_serialize_page_shift_count()); |
goetz@6458 | 2183 | |
goetz@6458 | 2184 | int mask = os::vm_page_size() - sizeof(int); |
goetz@6458 | 2185 | if (Assembler::is_simm(mask, 16)) { |
goetz@6458 | 2186 | andi(tmp2, tmp2, mask); |
goetz@6458 | 2187 | } else { |
goetz@6458 | 2188 | lis(tmp1, (int)((signed short) (mask >> 16))); |
goetz@6458 | 2189 | ori(tmp1, tmp1, mask & 0x0000ffff); |
goetz@6458 | 2190 | andr(tmp2, tmp2, tmp1); |
goetz@6458 | 2191 | } |
goetz@6458 | 2192 | |
goetz@6458 | 2193 | load_const(tmp1, (long) os::get_memory_serialize_page()); |
goetz@6458 | 2194 | release(); |
goetz@6458 | 2195 | stwx(R0, tmp1, tmp2); |
goetz@6458 | 2196 | } |
goetz@6458 | 2197 | |
goetz@6458 | 2198 | |
goetz@6458 | 2199 | // GC barrier helper macros |
goetz@6458 | 2200 | |
goetz@6458 | 2201 | // Write the card table byte if needed. |
goetz@6458 | 2202 | void MacroAssembler::card_write_barrier_post(Register Rstore_addr, Register Rnew_val, Register Rtmp) { |
goetz@6458 | 2203 | CardTableModRefBS* bs = (CardTableModRefBS*) Universe::heap()->barrier_set(); |
goetz@6458 | 2204 | assert(bs->kind() == BarrierSet::CardTableModRef || |
goetz@6458 | 2205 | bs->kind() == BarrierSet::CardTableExtension, "wrong barrier"); |
goetz@6458 | 2206 | #ifdef ASSERT |
goetz@6458 | 2207 | cmpdi(CCR0, Rnew_val, 0); |
goetz@6458 | 2208 | asm_assert_ne("null oop not allowed", 0x321); |
goetz@6458 | 2209 | #endif |
goetz@6458 | 2210 | card_table_write(bs->byte_map_base, Rtmp, Rstore_addr); |
goetz@6458 | 2211 | } |
goetz@6458 | 2212 | |
goetz@6458 | 2213 | // Write the card table byte. |
goetz@6458 | 2214 | void MacroAssembler::card_table_write(jbyte* byte_map_base, Register Rtmp, Register Robj) { |
goetz@6458 | 2215 | assert_different_registers(Robj, Rtmp, R0); |
goetz@6458 | 2216 | load_const_optimized(Rtmp, (address)byte_map_base, R0); |
goetz@6458 | 2217 | srdi(Robj, Robj, CardTableModRefBS::card_shift); |
goetz@6458 | 2218 | li(R0, 0); // dirty |
goetz@6501 | 2219 | if (UseConcMarkSweepGC) membar(Assembler::StoreStore); |
goetz@6458 | 2220 | stbx(R0, Rtmp, Robj); |
goetz@6458 | 2221 | } |
goetz@6458 | 2222 | |
phh@9669 | 2223 | // Kills R31 if value is a volatile register. |
phh@9669 | 2224 | void MacroAssembler::resolve_jobject(Register value, Register tmp1, Register tmp2, bool needs_frame) { |
phh@9669 | 2225 | Label done; |
phh@9669 | 2226 | cmpdi(CCR0, value, 0); |
phh@9669 | 2227 | beq(CCR0, done); // Use NULL as-is. |
phh@9669 | 2228 | |
phh@9669 | 2229 | clrrdi(tmp1, value, JNIHandles::weak_tag_size); |
phh@9669 | 2230 | #if INCLUDE_ALL_GCS |
phh@9669 | 2231 | if (UseG1GC) { andi_(tmp2, value, JNIHandles::weak_tag_mask); } |
phh@9669 | 2232 | #endif |
phh@9669 | 2233 | ld(value, 0, tmp1); // Resolve (untagged) jobject. |
phh@9669 | 2234 | |
phh@9669 | 2235 | #if INCLUDE_ALL_GCS |
phh@9669 | 2236 | if (UseG1GC) { |
phh@9669 | 2237 | Label not_weak; |
phh@9669 | 2238 | beq(CCR0, not_weak); // Test for jweak tag. |
phh@9669 | 2239 | verify_oop(value); |
phh@9669 | 2240 | g1_write_barrier_pre(noreg, // obj |
phh@9669 | 2241 | noreg, // offset |
phh@9669 | 2242 | value, // pre_val |
phh@9669 | 2243 | tmp1, tmp2, needs_frame); |
phh@9669 | 2244 | bind(not_weak); |
phh@9669 | 2245 | } |
phh@9669 | 2246 | #endif // INCLUDE_ALL_GCS |
phh@9669 | 2247 | verify_oop(value); |
phh@9669 | 2248 | bind(done); |
phh@9669 | 2249 | } |
phh@9669 | 2250 | |
goetz@6515 | 2251 | #if INCLUDE_ALL_GCS |
goetz@6458 | 2252 | // General G1 pre-barrier generator. |
goetz@6458 | 2253 | // Goal: record the previous value if it is not null. |
goetz@6458 | 2254 | void MacroAssembler::g1_write_barrier_pre(Register Robj, RegisterOrConstant offset, Register Rpre_val, |
goetz@6458 | 2255 | Register Rtmp1, Register Rtmp2, bool needs_frame) { |
goetz@6458 | 2256 | Label runtime, filtered; |
goetz@6458 | 2257 | |
goetz@6458 | 2258 | // Is marking active? |
goetz@6458 | 2259 | if (in_bytes(PtrQueue::byte_width_of_active()) == 4) { |
goetz@6458 | 2260 | lwz(Rtmp1, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()), R16_thread); |
goetz@6458 | 2261 | } else { |
goetz@6458 | 2262 | guarantee(in_bytes(PtrQueue::byte_width_of_active()) == 1, "Assumption"); |
goetz@6458 | 2263 | lbz(Rtmp1, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()), R16_thread); |
goetz@6458 | 2264 | } |
goetz@6458 | 2265 | cmpdi(CCR0, Rtmp1, 0); |
goetz@6458 | 2266 | beq(CCR0, filtered); |
goetz@6458 | 2267 | |
goetz@6458 | 2268 | // Do we need to load the previous value? |
goetz@6458 | 2269 | if (Robj != noreg) { |
goetz@6458 | 2270 | // Load the previous value... |
goetz@6458 | 2271 | if (UseCompressedOops) { |
goetz@6458 | 2272 | lwz(Rpre_val, offset, Robj); |
goetz@6458 | 2273 | } else { |
goetz@6458 | 2274 | ld(Rpre_val, offset, Robj); |
goetz@6458 | 2275 | } |
goetz@6458 | 2276 | // Previous value has been loaded into Rpre_val. |
goetz@6458 | 2277 | } |
goetz@6458 | 2278 | assert(Rpre_val != noreg, "must have a real register"); |
goetz@6458 | 2279 | |
goetz@6458 | 2280 | // Is the previous value null? |
goetz@6458 | 2281 | cmpdi(CCR0, Rpre_val, 0); |
goetz@6458 | 2282 | beq(CCR0, filtered); |
goetz@6458 | 2283 | |
goetz@6458 | 2284 | if (Robj != noreg && UseCompressedOops) { |
goetz@6458 | 2285 | decode_heap_oop_not_null(Rpre_val); |
goetz@6458 | 2286 | } |
goetz@6458 | 2287 | |
goetz@6458 | 2288 | // OK, it's not filtered, so we'll need to call enqueue. In the normal |
goetz@6458 | 2289 | // case, pre_val will be a scratch G-reg, but there are some cases in |
goetz@6458 | 2290 | // which it's an O-reg. In the first case, do a normal call. In the |
goetz@6458 | 2291 | // latter, do a save here and call the frameless version. |
goetz@6458 | 2292 | |
goetz@6458 | 2293 | // Can we store original value in the thread's buffer? |
goetz@6458 | 2294 | // Is index == 0? |
goetz@6458 | 2295 | // (The index field is typed as size_t.) |
goetz@6458 | 2296 | const Register Rbuffer = Rtmp1, Rindex = Rtmp2; |
goetz@6458 | 2297 | |
goetz@6458 | 2298 | ld(Rindex, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_index()), R16_thread); |
goetz@6458 | 2299 | cmpdi(CCR0, Rindex, 0); |
goetz@6458 | 2300 | beq(CCR0, runtime); // If index == 0, goto runtime. |
goetz@6458 | 2301 | ld(Rbuffer, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_buf()), R16_thread); |
goetz@6458 | 2302 | |
goetz@6458 | 2303 | addi(Rindex, Rindex, -wordSize); // Decrement index. |
goetz@6458 | 2304 | std(Rindex, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_index()), R16_thread); |
goetz@6458 | 2305 | |
goetz@6458 | 2306 | // Record the previous value. |
goetz@6458 | 2307 | stdx(Rpre_val, Rbuffer, Rindex); |
goetz@6458 | 2308 | b(filtered); |
goetz@6458 | 2309 | |
goetz@6458 | 2310 | bind(runtime); |
goetz@6458 | 2311 | |
phh@9669 | 2312 | // May need to preserve LR. Also needed if current frame is not compatible with C calling convention. |
goetz@6458 | 2313 | if (needs_frame) { |
goetz@6458 | 2314 | save_LR_CR(Rtmp1); |
goetz@6511 | 2315 | push_frame_reg_args(0, Rtmp2); |
goetz@6458 | 2316 | } |
goetz@6458 | 2317 | |
goetz@6458 | 2318 | if (Rpre_val->is_volatile() && Robj == noreg) mr(R31, Rpre_val); // Save pre_val across C call if it was preloaded. |
goetz@6458 | 2319 | call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), Rpre_val, R16_thread); |
goetz@6458 | 2320 | if (Rpre_val->is_volatile() && Robj == noreg) mr(Rpre_val, R31); // restore |
goetz@6458 | 2321 | |
goetz@6458 | 2322 | if (needs_frame) { |
goetz@6458 | 2323 | pop_frame(); |
goetz@6458 | 2324 | restore_LR_CR(Rtmp1); |
goetz@6458 | 2325 | } |
goetz@6458 | 2326 | |
goetz@6458 | 2327 | bind(filtered); |
goetz@6458 | 2328 | } |
goetz@6458 | 2329 | |
goetz@6458 | 2330 | // General G1 post-barrier generator |
goetz@6458 | 2331 | // Store cross-region card. |
goetz@6458 | 2332 | void MacroAssembler::g1_write_barrier_post(Register Rstore_addr, Register Rnew_val, Register Rtmp1, Register Rtmp2, Register Rtmp3, Label *filtered_ext) { |
goetz@6458 | 2333 | Label runtime, filtered_int; |
goetz@6458 | 2334 | Label& filtered = (filtered_ext != NULL) ? *filtered_ext : filtered_int; |
goetz@6458 | 2335 | assert_different_registers(Rstore_addr, Rnew_val, Rtmp1, Rtmp2); |
goetz@6458 | 2336 | |
goetz@6458 | 2337 | G1SATBCardTableModRefBS* bs = (G1SATBCardTableModRefBS*) Universe::heap()->barrier_set(); |
goetz@6458 | 2338 | assert(bs->kind() == BarrierSet::G1SATBCT || |
goetz@6458 | 2339 | bs->kind() == BarrierSet::G1SATBCTLogging, "wrong barrier"); |
goetz@6458 | 2340 | |
goetz@6458 | 2341 | // Does store cross heap regions? |
goetz@6458 | 2342 | if (G1RSBarrierRegionFilter) { |
goetz@6458 | 2343 | xorr(Rtmp1, Rstore_addr, Rnew_val); |
goetz@6458 | 2344 | srdi_(Rtmp1, Rtmp1, HeapRegion::LogOfHRGrainBytes); |
goetz@6458 | 2345 | beq(CCR0, filtered); |
goetz@6458 | 2346 | } |
goetz@6458 | 2347 | |
goetz@6458 | 2348 | // Crosses regions, storing NULL? |
goetz@6458 | 2349 | #ifdef ASSERT |
goetz@6458 | 2350 | cmpdi(CCR0, Rnew_val, 0); |
goetz@6458 | 2351 | asm_assert_ne("null oop not allowed (G1)", 0x322); // Checked by caller on PPC64, so following branch is obsolete: |
goetz@6458 | 2352 | //beq(CCR0, filtered); |
goetz@6458 | 2353 | #endif |
goetz@6458 | 2354 | |
goetz@6458 | 2355 | // Storing region crossing non-NULL, is card already dirty? |
goetz@6458 | 2356 | assert(sizeof(*bs->byte_map_base) == sizeof(jbyte), "adjust this code"); |
goetz@6458 | 2357 | const Register Rcard_addr = Rtmp1; |
goetz@6458 | 2358 | Register Rbase = Rtmp2; |
goetz@6458 | 2359 | load_const_optimized(Rbase, (address)bs->byte_map_base, /*temp*/ Rtmp3); |
goetz@6458 | 2360 | |
goetz@6458 | 2361 | srdi(Rcard_addr, Rstore_addr, CardTableModRefBS::card_shift); |
goetz@6458 | 2362 | |
goetz@6458 | 2363 | // Get the address of the card. |
goetz@6458 | 2364 | lbzx(/*card value*/ Rtmp3, Rbase, Rcard_addr); |
goetz@6515 | 2365 | cmpwi(CCR0, Rtmp3, (int)G1SATBCardTableModRefBS::g1_young_card_val()); |
goetz@6515 | 2366 | beq(CCR0, filtered); |
goetz@6515 | 2367 | |
goetz@6515 | 2368 | membar(Assembler::StoreLoad); |
goetz@6515 | 2369 | lbzx(/*card value*/ Rtmp3, Rbase, Rcard_addr); // Reload after membar. |
goetz@6515 | 2370 | cmpwi(CCR0, Rtmp3 /* card value */, CardTableModRefBS::dirty_card_val()); |
goetz@6458 | 2371 | beq(CCR0, filtered); |
goetz@6458 | 2372 | |
goetz@6458 | 2373 | // Storing a region crossing, non-NULL oop, card is clean. |
goetz@6458 | 2374 | // Dirty card and log. |
goetz@6515 | 2375 | li(Rtmp3, CardTableModRefBS::dirty_card_val()); |
goetz@6458 | 2376 | //release(); // G1: oops are allowed to get visible after dirty marking. |
goetz@6458 | 2377 | stbx(Rtmp3, Rbase, Rcard_addr); |
goetz@6458 | 2378 | |
goetz@6458 | 2379 | add(Rcard_addr, Rbase, Rcard_addr); // This is the address which needs to get enqueued. |
goetz@6458 | 2380 | Rbase = noreg; // end of lifetime |
goetz@6458 | 2381 | |
goetz@6458 | 2382 | const Register Rqueue_index = Rtmp2, |
goetz@6458 | 2383 | Rqueue_buf = Rtmp3; |
goetz@6458 | 2384 | ld(Rqueue_index, in_bytes(JavaThread::dirty_card_queue_offset() + PtrQueue::byte_offset_of_index()), R16_thread); |
goetz@6458 | 2385 | cmpdi(CCR0, Rqueue_index, 0); |
goetz@6458 | 2386 | beq(CCR0, runtime); // index == 0 then jump to runtime |
goetz@6458 | 2387 | ld(Rqueue_buf, in_bytes(JavaThread::dirty_card_queue_offset() + PtrQueue::byte_offset_of_buf()), R16_thread); |
goetz@6458 | 2388 | |
goetz@6458 | 2389 | addi(Rqueue_index, Rqueue_index, -wordSize); // decrement index |
goetz@6458 | 2390 | std(Rqueue_index, in_bytes(JavaThread::dirty_card_queue_offset() + PtrQueue::byte_offset_of_index()), R16_thread); |
goetz@6458 | 2391 | |
goetz@6458 | 2392 | stdx(Rcard_addr, Rqueue_buf, Rqueue_index); // store card |
goetz@6458 | 2393 | b(filtered); |
goetz@6458 | 2394 | |
goetz@6458 | 2395 | bind(runtime); |
goetz@6458 | 2396 | |
goetz@6458 | 2397 | // Save the live input values. |
goetz@6458 | 2398 | call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), Rcard_addr, R16_thread); |
goetz@6458 | 2399 | |
goetz@6458 | 2400 | bind(filtered_int); |
goetz@6458 | 2401 | } |
goetz@6515 | 2402 | #endif // INCLUDE_ALL_GCS |
goetz@6458 | 2403 | |
goetz@6458 | 2404 | // Values for last_Java_pc, and last_Java_sp must comply to the rules |
goetz@7222 | 2405 | // in frame_ppc.hpp. |
goetz@6458 | 2406 | void MacroAssembler::set_last_Java_frame(Register last_Java_sp, Register last_Java_pc) { |
goetz@6458 | 2407 | // Always set last_Java_pc and flags first because once last_Java_sp |
goetz@6458 | 2408 | // is visible has_last_Java_frame is true and users will look at the |
goetz@6458 | 2409 | // rest of the fields. (Note: flags should always be zero before we |
goetz@6458 | 2410 | // get here so doesn't need to be set.) |
goetz@6458 | 2411 | |
goetz@6458 | 2412 | // Verify that last_Java_pc was zeroed on return to Java |
goetz@6458 | 2413 | asm_assert_mem8_is_zero(in_bytes(JavaThread::last_Java_pc_offset()), R16_thread, |
goetz@6458 | 2414 | "last_Java_pc not zeroed before leaving Java", 0x200); |
goetz@6458 | 2415 | |
goetz@6458 | 2416 | // When returning from calling out from Java mode the frame anchor's |
goetz@6458 | 2417 | // last_Java_pc will always be set to NULL. It is set here so that |
goetz@6458 | 2418 | // if we are doing a call to native (not VM) that we capture the |
goetz@6458 | 2419 | // known pc and don't have to rely on the native call having a |
goetz@6458 | 2420 | // standard frame linkage where we can find the pc. |
goetz@6458 | 2421 | if (last_Java_pc != noreg) |
goetz@6458 | 2422 | std(last_Java_pc, in_bytes(JavaThread::last_Java_pc_offset()), R16_thread); |
goetz@6458 | 2423 | |
goetz@6495 | 2424 | // Set last_Java_sp last. |
goetz@6458 | 2425 | std(last_Java_sp, in_bytes(JavaThread::last_Java_sp_offset()), R16_thread); |
goetz@6458 | 2426 | } |
goetz@6458 | 2427 | |
goetz@6458 | 2428 | void MacroAssembler::reset_last_Java_frame(void) { |
goetz@6458 | 2429 | asm_assert_mem8_isnot_zero(in_bytes(JavaThread::last_Java_sp_offset()), |
goetz@6458 | 2430 | R16_thread, "SP was not set, still zero", 0x202); |
goetz@6458 | 2431 | |
goetz@6458 | 2432 | BLOCK_COMMENT("reset_last_Java_frame {"); |
goetz@6458 | 2433 | li(R0, 0); |
goetz@6458 | 2434 | |
goetz@6458 | 2435 | // _last_Java_sp = 0 |
goetz@6458 | 2436 | std(R0, in_bytes(JavaThread::last_Java_sp_offset()), R16_thread); |
goetz@6458 | 2437 | |
goetz@6458 | 2438 | // _last_Java_pc = 0 |
goetz@6458 | 2439 | std(R0, in_bytes(JavaThread::last_Java_pc_offset()), R16_thread); |
goetz@6458 | 2440 | BLOCK_COMMENT("} reset_last_Java_frame"); |
goetz@6458 | 2441 | } |
goetz@6458 | 2442 | |
goetz@6458 | 2443 | void MacroAssembler::set_top_ijava_frame_at_SP_as_last_Java_frame(Register sp, Register tmp1) { |
goetz@6458 | 2444 | assert_different_registers(sp, tmp1); |
goetz@6458 | 2445 | |
goetz@6458 | 2446 | // sp points to a TOP_IJAVA_FRAME, retrieve frame's PC via |
goetz@6458 | 2447 | // TOP_IJAVA_FRAME_ABI. |
goetz@6458 | 2448 | // FIXME: assert that we really have a TOP_IJAVA_FRAME here! |
goetz@6458 | 2449 | #ifdef CC_INTERP |
goetz@6458 | 2450 | ld(tmp1/*pc*/, _top_ijava_frame_abi(frame_manager_lr), sp); |
goetz@6458 | 2451 | #else |
goetz@6512 | 2452 | address entry = pc(); |
goetz@6512 | 2453 | load_const_optimized(tmp1, entry); |
goetz@6458 | 2454 | #endif |
goetz@6458 | 2455 | |
goetz@6458 | 2456 | set_last_Java_frame(/*sp=*/sp, /*pc=*/tmp1); |
goetz@6458 | 2457 | } |
goetz@6458 | 2458 | |
goetz@6458 | 2459 | void MacroAssembler::get_vm_result(Register oop_result) { |
goetz@6458 | 2460 | // Read: |
goetz@6458 | 2461 | // R16_thread |
goetz@6458 | 2462 | // R16_thread->in_bytes(JavaThread::vm_result_offset()) |
goetz@6458 | 2463 | // |
goetz@6458 | 2464 | // Updated: |
goetz@6458 | 2465 | // oop_result |
goetz@6458 | 2466 | // R16_thread->in_bytes(JavaThread::vm_result_offset()) |
goetz@6458 | 2467 | |
goetz@6458 | 2468 | ld(oop_result, in_bytes(JavaThread::vm_result_offset()), R16_thread); |
goetz@6458 | 2469 | li(R0, 0); |
goetz@6458 | 2470 | std(R0, in_bytes(JavaThread::vm_result_offset()), R16_thread); |
goetz@6458 | 2471 | |
goetz@6458 | 2472 | verify_oop(oop_result); |
goetz@6458 | 2473 | } |
goetz@6458 | 2474 | |
goetz@6458 | 2475 | void MacroAssembler::get_vm_result_2(Register metadata_result) { |
goetz@6458 | 2476 | // Read: |
goetz@6458 | 2477 | // R16_thread |
goetz@6458 | 2478 | // R16_thread->in_bytes(JavaThread::vm_result_2_offset()) |
goetz@6458 | 2479 | // |
goetz@6458 | 2480 | // Updated: |
goetz@6458 | 2481 | // metadata_result |
goetz@6458 | 2482 | // R16_thread->in_bytes(JavaThread::vm_result_2_offset()) |
goetz@6458 | 2483 | |
goetz@6458 | 2484 | ld(metadata_result, in_bytes(JavaThread::vm_result_2_offset()), R16_thread); |
goetz@6458 | 2485 | li(R0, 0); |
goetz@6458 | 2486 | std(R0, in_bytes(JavaThread::vm_result_2_offset()), R16_thread); |
goetz@6458 | 2487 | } |
goetz@6458 | 2488 | |
goetz@6458 | 2489 | |
goetz@6458 | 2490 | void MacroAssembler::encode_klass_not_null(Register dst, Register src) { |
goetz@6501 | 2491 | Register current = (src != noreg) ? src : dst; // Klass is in dst if no src provided. |
goetz@6477 | 2492 | if (Universe::narrow_klass_base() != 0) { |
goetz@6515 | 2493 | // Use dst as temp if it is free. |
goetz@6515 | 2494 | load_const(R0, Universe::narrow_klass_base(), (dst != current && dst != R0) ? dst : noreg); |
goetz@6501 | 2495 | sub(dst, current, R0); |
goetz@6501 | 2496 | current = dst; |
goetz@6477 | 2497 | } |
goetz@6501 | 2498 | if (Universe::narrow_klass_shift() != 0) { |
goetz@6501 | 2499 | srdi(dst, current, Universe::narrow_klass_shift()); |
goetz@6501 | 2500 | current = dst; |
goetz@6458 | 2501 | } |
goetz@6501 | 2502 | mr_if_needed(dst, current); // Move may be required. |
goetz@6458 | 2503 | } |
goetz@6458 | 2504 | |
goetz@6458 | 2505 | void MacroAssembler::store_klass(Register dst_oop, Register klass, Register ck) { |
goetz@6474 | 2506 | if (UseCompressedClassPointers) { |
goetz@6458 | 2507 | encode_klass_not_null(ck, klass); |
goetz@6458 | 2508 | stw(ck, oopDesc::klass_offset_in_bytes(), dst_oop); |
goetz@6458 | 2509 | } else { |
goetz@6458 | 2510 | std(klass, oopDesc::klass_offset_in_bytes(), dst_oop); |
goetz@6458 | 2511 | } |
goetz@6458 | 2512 | } |
goetz@6458 | 2513 | |
goetz@6512 | 2514 | void MacroAssembler::store_klass_gap(Register dst_oop, Register val) { |
goetz@6512 | 2515 | if (UseCompressedClassPointers) { |
goetz@6512 | 2516 | if (val == noreg) { |
goetz@6512 | 2517 | val = R0; |
goetz@6512 | 2518 | li(val, 0); |
goetz@6512 | 2519 | } |
goetz@6512 | 2520 | stw(val, oopDesc::klass_gap_offset_in_bytes(), dst_oop); // klass gap if compressed |
goetz@6512 | 2521 | } |
goetz@6512 | 2522 | } |
goetz@6512 | 2523 | |
goetz@6477 | 2524 | int MacroAssembler::instr_size_for_decode_klass_not_null() { |
goetz@6477 | 2525 | if (!UseCompressedClassPointers) return 0; |
goetz@6477 | 2526 | int num_instrs = 1; // shift or move |
goetz@6477 | 2527 | if (Universe::narrow_klass_base() != 0) num_instrs = 7; // shift + load const + add |
goetz@6477 | 2528 | return num_instrs * BytesPerInstWord; |
goetz@6477 | 2529 | } |
goetz@6477 | 2530 | |
goetz@6458 | 2531 | void MacroAssembler::decode_klass_not_null(Register dst, Register src) { |
goetz@7222 | 2532 | assert(dst != R0, "Dst reg may not be R0, as R0 is used here."); |
goetz@6458 | 2533 | if (src == noreg) src = dst; |
goetz@6463 | 2534 | Register shifted_src = src; |
goetz@6477 | 2535 | if (Universe::narrow_klass_shift() != 0 || |
goetz@6477 | 2536 | Universe::narrow_klass_base() == 0 && src != dst) { // Move required. |
goetz@6463 | 2537 | shifted_src = dst; |
goetz@6463 | 2538 | sldi(shifted_src, src, Universe::narrow_klass_shift()); |
goetz@6458 | 2539 | } |
goetz@6477 | 2540 | if (Universe::narrow_klass_base() != 0) { |
goetz@6477 | 2541 | load_const(R0, Universe::narrow_klass_base()); |
goetz@6477 | 2542 | add(dst, shifted_src, R0); |
goetz@6477 | 2543 | } |
goetz@6458 | 2544 | } |
goetz@6458 | 2545 | |
goetz@6458 | 2546 | void MacroAssembler::load_klass(Register dst, Register src) { |
goetz@6474 | 2547 | if (UseCompressedClassPointers) { |
goetz@6458 | 2548 | lwz(dst, oopDesc::klass_offset_in_bytes(), src); |
goetz@6458 | 2549 | // Attention: no null check here! |
goetz@6458 | 2550 | decode_klass_not_null(dst, dst); |
goetz@6458 | 2551 | } else { |
goetz@6458 | 2552 | ld(dst, oopDesc::klass_offset_in_bytes(), src); |
goetz@6458 | 2553 | } |
goetz@6458 | 2554 | } |
goetz@6458 | 2555 | |
goetz@6458 | 2556 | void MacroAssembler::load_klass_with_trap_null_check(Register dst, Register src) { |
goetz@6486 | 2557 | if (!os::zero_page_read_protected()) { |
goetz@6458 | 2558 | if (TrapBasedNullChecks) { |
goetz@6458 | 2559 | trap_null_check(src); |
goetz@6458 | 2560 | } |
goetz@6458 | 2561 | } |
goetz@6458 | 2562 | load_klass(dst, src); |
goetz@6458 | 2563 | } |
goetz@6458 | 2564 | |
goetz@6458 | 2565 | void MacroAssembler::reinit_heapbase(Register d, Register tmp) { |
goetz@6463 | 2566 | if (Universe::heap() != NULL) { |
goetz@7222 | 2567 | load_const_optimized(R30, Universe::narrow_ptrs_base(), tmp); |
goetz@6463 | 2568 | } else { |
goetz@7222 | 2569 | // Heap not yet allocated. Load indirectly. |
goetz@7222 | 2570 | int simm16_offset = load_const_optimized(R30, Universe::narrow_ptrs_base_addr(), tmp, true); |
goetz@7222 | 2571 | ld(R30, simm16_offset, R30); |
goetz@6458 | 2572 | } |
goetz@6458 | 2573 | } |
goetz@6458 | 2574 | |
goetz@6495 | 2575 | // Clear Array |
goetz@6495 | 2576 | // Kills both input registers. tmp == R0 is allowed. |
goetz@6495 | 2577 | void MacroAssembler::clear_memory_doubleword(Register base_ptr, Register cnt_dwords, Register tmp) { |
goetz@6495 | 2578 | // Procedure for large arrays (uses data cache block zero instruction). |
goetz@6495 | 2579 | Label startloop, fast, fastloop, small_rest, restloop, done; |
goetz@6495 | 2580 | const int cl_size = VM_Version::get_cache_line_size(), |
goetz@6495 | 2581 | cl_dwords = cl_size>>3, |
goetz@6495 | 2582 | cl_dw_addr_bits = exact_log2(cl_dwords), |
goetz@6495 | 2583 | dcbz_min = 1; // Min count of dcbz executions, needs to be >0. |
goetz@6495 | 2584 | |
goetz@6495 | 2585 | //2: |
goetz@6495 | 2586 | cmpdi(CCR1, cnt_dwords, ((dcbz_min+1)<<cl_dw_addr_bits)-1); // Big enough? (ensure >=dcbz_min lines included). |
goetz@6495 | 2587 | blt(CCR1, small_rest); // Too small. |
goetz@6495 | 2588 | rldicl_(tmp, base_ptr, 64-3, 64-cl_dw_addr_bits); // Extract dword offset within first cache line. |
goetz@6495 | 2589 | beq(CCR0, fast); // Already 128byte aligned. |
goetz@6495 | 2590 | |
goetz@6495 | 2591 | subfic(tmp, tmp, cl_dwords); |
goetz@6495 | 2592 | mtctr(tmp); // Set ctr to hit 128byte boundary (0<ctr<cl_dwords). |
goetz@6495 | 2593 | subf(cnt_dwords, tmp, cnt_dwords); // rest. |
goetz@6495 | 2594 | li(tmp, 0); |
goetz@6495 | 2595 | //10: |
goetz@6495 | 2596 | bind(startloop); // Clear at the beginning to reach 128byte boundary. |
goetz@6495 | 2597 | std(tmp, 0, base_ptr); // Clear 8byte aligned block. |
goetz@6495 | 2598 | addi(base_ptr, base_ptr, 8); |
goetz@6495 | 2599 | bdnz(startloop); |
goetz@6495 | 2600 | //13: |
goetz@6495 | 2601 | bind(fast); // Clear 128byte blocks. |
goetz@6495 | 2602 | srdi(tmp, cnt_dwords, cl_dw_addr_bits); // Loop count for 128byte loop (>0). |
goetz@6495 | 2603 | andi(cnt_dwords, cnt_dwords, cl_dwords-1); // Rest in dwords. |
goetz@6495 | 2604 | mtctr(tmp); // Load counter. |
goetz@6495 | 2605 | //16: |
goetz@6495 | 2606 | bind(fastloop); |
goetz@6495 | 2607 | dcbz(base_ptr); // Clear 128byte aligned block. |
goetz@6495 | 2608 | addi(base_ptr, base_ptr, cl_size); |
goetz@6495 | 2609 | bdnz(fastloop); |
goetz@6495 | 2610 | if (InsertEndGroupPPC64) { endgroup(); } else { nop(); } |
goetz@6495 | 2611 | //20: |
goetz@6495 | 2612 | bind(small_rest); |
goetz@6495 | 2613 | cmpdi(CCR0, cnt_dwords, 0); // size 0? |
goetz@6495 | 2614 | beq(CCR0, done); // rest == 0 |
goetz@6495 | 2615 | li(tmp, 0); |
goetz@6495 | 2616 | mtctr(cnt_dwords); // Load counter. |
goetz@6495 | 2617 | //24: |
goetz@6495 | 2618 | bind(restloop); // Clear rest. |
goetz@6495 | 2619 | std(tmp, 0, base_ptr); // Clear 8byte aligned block. |
goetz@6495 | 2620 | addi(base_ptr, base_ptr, 8); |
goetz@6495 | 2621 | bdnz(restloop); |
goetz@6495 | 2622 | //27: |
goetz@6495 | 2623 | bind(done); |
goetz@6495 | 2624 | } |
goetz@6495 | 2625 | |
goetz@6458 | 2626 | /////////////////////////////////////////// String intrinsics //////////////////////////////////////////// |
goetz@6458 | 2627 | |
goetz@6458 | 2628 | // Search for a single jchar in an jchar[]. |
goetz@6458 | 2629 | // |
goetz@6458 | 2630 | // Assumes that result differs from all other registers. |
goetz@6458 | 2631 | // |
goetz@6458 | 2632 | // Haystack, needle are the addresses of jchar-arrays. |
goetz@6458 | 2633 | // NeedleChar is needle[0] if it is known at compile time. |
goetz@6458 | 2634 | // Haycnt is the length of the haystack. We assume haycnt >=1. |
goetz@6458 | 2635 | // |
goetz@6458 | 2636 | // Preserves haystack, haycnt, kills all other registers. |
goetz@6458 | 2637 | // |
goetz@6458 | 2638 | // If needle == R0, we search for the constant needleChar. |
goetz@6458 | 2639 | void MacroAssembler::string_indexof_1(Register result, Register haystack, Register haycnt, |
goetz@6458 | 2640 | Register needle, jchar needleChar, |
goetz@6458 | 2641 | Register tmp1, Register tmp2) { |
goetz@6458 | 2642 | |
goetz@6458 | 2643 | assert_different_registers(result, haystack, haycnt, needle, tmp1, tmp2); |
goetz@6458 | 2644 | |
goetz@6458 | 2645 | Label L_InnerLoop, L_FinalCheck, L_Found1, L_Found2, L_Found3, L_NotFound, L_End; |
goetz@6458 | 2646 | Register needle0 = needle, // Contains needle[0]. |
goetz@6458 | 2647 | addr = tmp1, |
goetz@6458 | 2648 | ch1 = tmp2, |
goetz@6458 | 2649 | ch2 = R0; |
goetz@6458 | 2650 | |
goetz@6458 | 2651 | //2 (variable) or 3 (const): |
goetz@6458 | 2652 | if (needle != R0) lhz(needle0, 0, needle); // Preload needle character, needle has len==1. |
goetz@6458 | 2653 | dcbtct(haystack, 0x00); // Indicate R/O access to haystack. |
goetz@6458 | 2654 | |
goetz@6458 | 2655 | srwi_(tmp2, haycnt, 1); // Shift right by exact_log2(UNROLL_FACTOR). |
goetz@6458 | 2656 | mr(addr, haystack); |
goetz@6458 | 2657 | beq(CCR0, L_FinalCheck); |
goetz@6458 | 2658 | mtctr(tmp2); // Move to count register. |
goetz@6458 | 2659 | //8: |
goetz@6458 | 2660 | bind(L_InnerLoop); // Main work horse (2x unrolled search loop). |
goetz@6458 | 2661 | lhz(ch1, 0, addr); // Load characters from haystack. |
goetz@6458 | 2662 | lhz(ch2, 2, addr); |
goetz@6458 | 2663 | (needle != R0) ? cmpw(CCR0, ch1, needle0) : cmplwi(CCR0, ch1, needleChar); |
goetz@6458 | 2664 | (needle != R0) ? cmpw(CCR1, ch2, needle0) : cmplwi(CCR1, ch2, needleChar); |
goetz@6458 | 2665 | beq(CCR0, L_Found1); // Did we find the needle? |
goetz@6458 | 2666 | beq(CCR1, L_Found2); |
goetz@6458 | 2667 | addi(addr, addr, 4); |
goetz@6458 | 2668 | bdnz(L_InnerLoop); |
goetz@6458 | 2669 | //16: |
goetz@6458 | 2670 | bind(L_FinalCheck); |
goetz@6458 | 2671 | andi_(R0, haycnt, 1); |
goetz@6458 | 2672 | beq(CCR0, L_NotFound); |
goetz@6458 | 2673 | lhz(ch1, 0, addr); // One position left at which we have to compare. |
goetz@6458 | 2674 | (needle != R0) ? cmpw(CCR1, ch1, needle0) : cmplwi(CCR1, ch1, needleChar); |
goetz@6458 | 2675 | beq(CCR1, L_Found3); |
goetz@6458 | 2676 | //21: |
goetz@6458 | 2677 | bind(L_NotFound); |
goetz@6458 | 2678 | li(result, -1); // Not found. |
goetz@6458 | 2679 | b(L_End); |
goetz@6458 | 2680 | |
goetz@6458 | 2681 | bind(L_Found2); |
goetz@6458 | 2682 | addi(addr, addr, 2); |
goetz@6458 | 2683 | //24: |
goetz@6458 | 2684 | bind(L_Found1); |
goetz@6458 | 2685 | bind(L_Found3); // Return index ... |
goetz@6458 | 2686 | subf(addr, haystack, addr); // relative to haystack, |
goetz@6458 | 2687 | srdi(result, addr, 1); // in characters. |
goetz@6458 | 2688 | bind(L_End); |
goetz@6458 | 2689 | } |
goetz@6458 | 2690 | |
goetz@6458 | 2691 | |
goetz@6458 | 2692 | // Implementation of IndexOf for jchar arrays. |
goetz@6458 | 2693 | // |
goetz@6458 | 2694 | // The length of haystack and needle are not constant, i.e. passed in a register. |
goetz@6458 | 2695 | // |
goetz@6458 | 2696 | // Preserves registers haystack, needle. |
goetz@6458 | 2697 | // Kills registers haycnt, needlecnt. |
goetz@6458 | 2698 | // Assumes that result differs from all other registers. |
goetz@6458 | 2699 | // Haystack, needle are the addresses of jchar-arrays. |
goetz@6458 | 2700 | // Haycnt, needlecnt are the lengths of them, respectively. |
goetz@6458 | 2701 | // |
goetz@6458 | 2702 | // Needlecntval must be zero or 15-bit unsigned immediate and > 1. |
goetz@6458 | 2703 | void MacroAssembler::string_indexof(Register result, Register haystack, Register haycnt, |
goetz@6458 | 2704 | Register needle, ciTypeArray* needle_values, Register needlecnt, int needlecntval, |
goetz@6458 | 2705 | Register tmp1, Register tmp2, Register tmp3, Register tmp4) { |
goetz@6458 | 2706 | |
goetz@6458 | 2707 | // Ensure 0<needlecnt<=haycnt in ideal graph as prerequisite! |
goetz@6458 | 2708 | Label L_TooShort, L_Found, L_NotFound, L_End; |
goetz@6458 | 2709 | Register last_addr = haycnt, // Kill haycnt at the beginning. |
goetz@6458 | 2710 | addr = tmp1, |
goetz@6458 | 2711 | n_start = tmp2, |
goetz@6458 | 2712 | ch1 = tmp3, |
goetz@6458 | 2713 | ch2 = R0; |
goetz@6458 | 2714 | |
goetz@6458 | 2715 | // ************************************************************************************************** |
goetz@6458 | 2716 | // Prepare for main loop: optimized for needle count >=2, bail out otherwise. |
goetz@6458 | 2717 | // ************************************************************************************************** |
goetz@6458 | 2718 | |
goetz@6458 | 2719 | //1 (variable) or 3 (const): |
goetz@6458 | 2720 | dcbtct(needle, 0x00); // Indicate R/O access to str1. |
goetz@6458 | 2721 | dcbtct(haystack, 0x00); // Indicate R/O access to str2. |
goetz@6458 | 2722 | |
goetz@6458 | 2723 | // Compute last haystack addr to use if no match gets found. |
goetz@6458 | 2724 | if (needlecntval == 0) { // variable needlecnt |
goetz@6458 | 2725 | //3: |
goetz@6458 | 2726 | subf(ch1, needlecnt, haycnt); // Last character index to compare is haycnt-needlecnt. |
goetz@6458 | 2727 | addi(addr, haystack, -2); // Accesses use pre-increment. |
goetz@6458 | 2728 | cmpwi(CCR6, needlecnt, 2); |
goetz@6458 | 2729 | blt(CCR6, L_TooShort); // Variable needlecnt: handle short needle separately. |
goetz@6458 | 2730 | slwi(ch1, ch1, 1); // Scale to number of bytes. |
goetz@6458 | 2731 | lwz(n_start, 0, needle); // Load first 2 characters of needle. |
goetz@6458 | 2732 | add(last_addr, haystack, ch1); // Point to last address to compare (haystack+2*(haycnt-needlecnt)). |
goetz@6458 | 2733 | addi(needlecnt, needlecnt, -2); // Rest of needle. |
goetz@6458 | 2734 | } else { // constant needlecnt |
goetz@6458 | 2735 | guarantee(needlecntval != 1, "IndexOf with single-character needle must be handled separately"); |
goetz@6458 | 2736 | assert((needlecntval & 0x7fff) == needlecntval, "wrong immediate"); |
goetz@6458 | 2737 | //5: |
goetz@6458 | 2738 | addi(ch1, haycnt, -needlecntval); // Last character index to compare is haycnt-needlecnt. |
goetz@6458 | 2739 | lwz(n_start, 0, needle); // Load first 2 characters of needle. |
goetz@6458 | 2740 | addi(addr, haystack, -2); // Accesses use pre-increment. |
goetz@6458 | 2741 | slwi(ch1, ch1, 1); // Scale to number of bytes. |
goetz@6458 | 2742 | add(last_addr, haystack, ch1); // Point to last address to compare (haystack+2*(haycnt-needlecnt)). |
goetz@6458 | 2743 | li(needlecnt, needlecntval-2); // Rest of needle. |
goetz@6458 | 2744 | } |
goetz@6458 | 2745 | |
goetz@6458 | 2746 | // Main Loop (now we have at least 3 characters). |
goetz@6458 | 2747 | //11: |
goetz@6458 | 2748 | Label L_OuterLoop, L_InnerLoop, L_FinalCheck, L_Comp1, L_Comp2, L_Comp3; |
goetz@6458 | 2749 | bind(L_OuterLoop); // Search for 1st 2 characters. |
goetz@6458 | 2750 | Register addr_diff = tmp4; |
goetz@6458 | 2751 | subf(addr_diff, addr, last_addr); // Difference between already checked address and last address to check. |
goetz@6458 | 2752 | addi(addr, addr, 2); // This is the new address we want to use for comparing. |
goetz@6458 | 2753 | srdi_(ch2, addr_diff, 2); |
goetz@6458 | 2754 | beq(CCR0, L_FinalCheck); // 2 characters left? |
goetz@6458 | 2755 | mtctr(ch2); // addr_diff/4 |
goetz@6458 | 2756 | //16: |
goetz@6458 | 2757 | bind(L_InnerLoop); // Main work horse (2x unrolled search loop) |
goetz@6458 | 2758 | lwz(ch1, 0, addr); // Load 2 characters of haystack (ignore alignment). |
goetz@6458 | 2759 | lwz(ch2, 2, addr); |
goetz@6458 | 2760 | cmpw(CCR0, ch1, n_start); // Compare 2 characters (1 would be sufficient but try to reduce branches to CompLoop). |
goetz@6458 | 2761 | cmpw(CCR1, ch2, n_start); |
goetz@6458 | 2762 | beq(CCR0, L_Comp1); // Did we find the needle start? |
goetz@6458 | 2763 | beq(CCR1, L_Comp2); |
goetz@6458 | 2764 | addi(addr, addr, 4); |
goetz@6458 | 2765 | bdnz(L_InnerLoop); |
goetz@6458 | 2766 | //24: |
goetz@6458 | 2767 | bind(L_FinalCheck); |
goetz@6458 | 2768 | rldicl_(addr_diff, addr_diff, 64-1, 63); // Remaining characters not covered by InnerLoop: (addr_diff>>1)&1. |
goetz@6458 | 2769 | beq(CCR0, L_NotFound); |
goetz@6458 | 2770 | lwz(ch1, 0, addr); // One position left at which we have to compare. |
goetz@6458 | 2771 | cmpw(CCR1, ch1, n_start); |
goetz@6458 | 2772 | beq(CCR1, L_Comp3); |
goetz@6458 | 2773 | //29: |
goetz@6458 | 2774 | bind(L_NotFound); |
goetz@6458 | 2775 | li(result, -1); // not found |
goetz@6458 | 2776 | b(L_End); |
goetz@6458 | 2777 | |
goetz@6458 | 2778 | |
goetz@6458 | 2779 | // ************************************************************************************************** |
goetz@6458 | 2780 | // Special Case: unfortunately, the variable needle case can be called with needlecnt<2 |
goetz@6458 | 2781 | // ************************************************************************************************** |
goetz@6458 | 2782 | //31: |
goetz@6458 | 2783 | if ((needlecntval>>1) !=1 ) { // Const needlecnt is 2 or 3? Reduce code size. |
goetz@6458 | 2784 | int nopcnt = 5; |
goetz@6458 | 2785 | if (needlecntval !=0 ) ++nopcnt; // Balance alignment (other case: see below). |
goetz@6458 | 2786 | if (needlecntval == 0) { // We have to handle these cases separately. |
goetz@6458 | 2787 | Label L_OneCharLoop; |
goetz@6458 | 2788 | bind(L_TooShort); |
goetz@6458 | 2789 | mtctr(haycnt); |
goetz@6458 | 2790 | lhz(n_start, 0, needle); // First character of needle |
goetz@6458 | 2791 | bind(L_OneCharLoop); |
goetz@6458 | 2792 | lhzu(ch1, 2, addr); |
goetz@6458 | 2793 | cmpw(CCR1, ch1, n_start); |
goetz@6458 | 2794 | beq(CCR1, L_Found); // Did we find the one character needle? |
goetz@6458 | 2795 | bdnz(L_OneCharLoop); |
goetz@6458 | 2796 | li(result, -1); // Not found. |
goetz@6458 | 2797 | b(L_End); |
goetz@6458 | 2798 | } // 8 instructions, so no impact on alignment. |
goetz@6458 | 2799 | for (int x = 0; x < nopcnt; ++x) nop(); |
goetz@6458 | 2800 | } |
goetz@6458 | 2801 | |
goetz@6458 | 2802 | // ************************************************************************************************** |
goetz@6458 | 2803 | // Regular Case Part II: compare rest of needle (first 2 characters have been compared already) |
goetz@6458 | 2804 | // ************************************************************************************************** |
goetz@6458 | 2805 | |
goetz@6458 | 2806 | // Compare the rest |
goetz@6458 | 2807 | //36 if needlecntval==0, else 37: |
goetz@6458 | 2808 | bind(L_Comp2); |
goetz@6458 | 2809 | addi(addr, addr, 2); // First comparison has failed, 2nd one hit. |
goetz@6458 | 2810 | bind(L_Comp1); // Addr points to possible needle start. |
goetz@6458 | 2811 | bind(L_Comp3); // Could have created a copy and use a different return address but saving code size here. |
goetz@6458 | 2812 | if (needlecntval != 2) { // Const needlecnt==2? |
goetz@6458 | 2813 | if (needlecntval != 3) { |
goetz@6458 | 2814 | if (needlecntval == 0) beq(CCR6, L_Found); // Variable needlecnt==2? |
goetz@6458 | 2815 | Register ind_reg = tmp4; |
goetz@6458 | 2816 | li(ind_reg, 2*2); // First 2 characters are already compared, use index 2. |
goetz@6458 | 2817 | mtctr(needlecnt); // Decremented by 2, still > 0. |
goetz@6458 | 2818 | //40: |
goetz@6458 | 2819 | Label L_CompLoop; |
goetz@6458 | 2820 | bind(L_CompLoop); |
goetz@6458 | 2821 | lhzx(ch2, needle, ind_reg); |
goetz@6458 | 2822 | lhzx(ch1, addr, ind_reg); |
goetz@6458 | 2823 | cmpw(CCR1, ch1, ch2); |
goetz@6458 | 2824 | bne(CCR1, L_OuterLoop); |
goetz@6458 | 2825 | addi(ind_reg, ind_reg, 2); |
goetz@6458 | 2826 | bdnz(L_CompLoop); |
goetz@6458 | 2827 | } else { // No loop required if there's only one needle character left. |
goetz@6458 | 2828 | lhz(ch2, 2*2, needle); |
goetz@6458 | 2829 | lhz(ch1, 2*2, addr); |
goetz@6458 | 2830 | cmpw(CCR1, ch1, ch2); |
goetz@6458 | 2831 | bne(CCR1, L_OuterLoop); |
goetz@6458 | 2832 | } |
goetz@6458 | 2833 | } |
goetz@6458 | 2834 | // Return index ... |
goetz@6458 | 2835 | //46: |
goetz@6458 | 2836 | bind(L_Found); |
goetz@6458 | 2837 | subf(addr, haystack, addr); // relative to haystack, ... |
goetz@6458 | 2838 | srdi(result, addr, 1); // in characters. |
goetz@6458 | 2839 | //48: |
goetz@6458 | 2840 | bind(L_End); |
goetz@6458 | 2841 | } |
goetz@6458 | 2842 | |
goetz@6458 | 2843 | // Implementation of Compare for jchar arrays. |
goetz@6458 | 2844 | // |
goetz@6458 | 2845 | // Kills the registers str1, str2, cnt1, cnt2. |
goetz@6458 | 2846 | // Kills cr0, ctr. |
goetz@6458 | 2847 | // Assumes that result differes from the input registers. |
goetz@6458 | 2848 | void MacroAssembler::string_compare(Register str1_reg, Register str2_reg, Register cnt1_reg, Register cnt2_reg, |
goetz@6458 | 2849 | Register result_reg, Register tmp_reg) { |
goetz@6458 | 2850 | assert_different_registers(result_reg, str1_reg, str2_reg, cnt1_reg, cnt2_reg, tmp_reg); |
goetz@6458 | 2851 | |
goetz@6458 | 2852 | Label Ldone, Lslow_case, Lslow_loop, Lfast_loop; |
goetz@6458 | 2853 | Register cnt_diff = R0, |
goetz@6458 | 2854 | limit_reg = cnt1_reg, |
goetz@6458 | 2855 | chr1_reg = result_reg, |
goetz@6458 | 2856 | chr2_reg = cnt2_reg, |
goetz@6458 | 2857 | addr_diff = str2_reg; |
goetz@6458 | 2858 | |
goetz@6458 | 2859 | // Offset 0 should be 32 byte aligned. |
goetz@6458 | 2860 | //-4: |
goetz@6458 | 2861 | dcbtct(str1_reg, 0x00); // Indicate R/O access to str1. |
goetz@6458 | 2862 | dcbtct(str2_reg, 0x00); // Indicate R/O access to str2. |
goetz@6458 | 2863 | //-2: |
goetz@6458 | 2864 | // Compute min(cnt1, cnt2) and check if 0 (bail out if we don't need to compare characters). |
goetz@6458 | 2865 | subf(result_reg, cnt2_reg, cnt1_reg); // difference between cnt1/2 |
goetz@6458 | 2866 | subf_(addr_diff, str1_reg, str2_reg); // alias? |
goetz@6458 | 2867 | beq(CCR0, Ldone); // return cnt difference if both ones are identical |
goetz@6458 | 2868 | srawi(limit_reg, result_reg, 31); // generate signmask (cnt1/2 must be non-negative so cnt_diff can't overflow) |
goetz@6458 | 2869 | mr(cnt_diff, result_reg); |
goetz@6458 | 2870 | andr(limit_reg, result_reg, limit_reg); // difference or zero (negative): cnt1<cnt2 ? cnt1-cnt2 : 0 |
goetz@6458 | 2871 | add_(limit_reg, cnt2_reg, limit_reg); // min(cnt1, cnt2)==0? |
goetz@6458 | 2872 | beq(CCR0, Ldone); // return cnt difference if one has 0 length |
goetz@6458 | 2873 | |
goetz@6458 | 2874 | lhz(chr1_reg, 0, str1_reg); // optional: early out if first characters mismatch |
goetz@6458 | 2875 | lhzx(chr2_reg, str1_reg, addr_diff); // optional: early out if first characters mismatch |
goetz@6458 | 2876 | addi(tmp_reg, limit_reg, -1); // min(cnt1, cnt2)-1 |
goetz@6458 | 2877 | subf_(result_reg, chr2_reg, chr1_reg); // optional: early out if first characters mismatch |
goetz@6458 | 2878 | bne(CCR0, Ldone); // optional: early out if first characters mismatch |
goetz@6458 | 2879 | |
goetz@6458 | 2880 | // Set loop counter by scaling down tmp_reg |
goetz@6458 | 2881 | srawi_(chr2_reg, tmp_reg, exact_log2(4)); // (min(cnt1, cnt2)-1)/4 |
goetz@6458 | 2882 | ble(CCR0, Lslow_case); // need >4 characters for fast loop |
goetz@6458 | 2883 | andi(limit_reg, tmp_reg, 4-1); // remaining characters |
goetz@6458 | 2884 | |
goetz@6458 | 2885 | // Adapt str1_reg str2_reg for the first loop iteration |
goetz@6458 | 2886 | mtctr(chr2_reg); // (min(cnt1, cnt2)-1)/4 |
goetz@6458 | 2887 | addi(limit_reg, limit_reg, 4+1); // compare last 5-8 characters in slow_case if mismatch found in fast_loop |
goetz@6458 | 2888 | //16: |
goetz@6458 | 2889 | // Compare the rest of the characters |
goetz@6458 | 2890 | bind(Lfast_loop); |
goetz@6458 | 2891 | ld(chr1_reg, 0, str1_reg); |
goetz@6458 | 2892 | ldx(chr2_reg, str1_reg, addr_diff); |
goetz@6458 | 2893 | cmpd(CCR0, chr2_reg, chr1_reg); |
goetz@6458 | 2894 | bne(CCR0, Lslow_case); // return chr1_reg |
goetz@6458 | 2895 | addi(str1_reg, str1_reg, 4*2); |
goetz@6458 | 2896 | bdnz(Lfast_loop); |
goetz@6458 | 2897 | addi(limit_reg, limit_reg, -4); // no mismatch found in fast_loop, only 1-4 characters missing |
goetz@6458 | 2898 | //23: |
goetz@6458 | 2899 | bind(Lslow_case); |
goetz@6458 | 2900 | mtctr(limit_reg); |
goetz@6458 | 2901 | //24: |
goetz@6458 | 2902 | bind(Lslow_loop); |
goetz@6458 | 2903 | lhz(chr1_reg, 0, str1_reg); |
goetz@6458 | 2904 | lhzx(chr2_reg, str1_reg, addr_diff); |
goetz@6458 | 2905 | subf_(result_reg, chr2_reg, chr1_reg); |
goetz@6458 | 2906 | bne(CCR0, Ldone); // return chr1_reg |
goetz@6458 | 2907 | addi(str1_reg, str1_reg, 1*2); |
goetz@6458 | 2908 | bdnz(Lslow_loop); |
goetz@6458 | 2909 | //30: |
goetz@6458 | 2910 | // If strings are equal up to min length, return the length difference. |
goetz@6458 | 2911 | mr(result_reg, cnt_diff); |
goetz@6458 | 2912 | nop(); // alignment |
goetz@6458 | 2913 | //32: |
goetz@6458 | 2914 | // Otherwise, return the difference between the first mismatched chars. |
goetz@6458 | 2915 | bind(Ldone); |
goetz@6458 | 2916 | } |
goetz@6458 | 2917 | |
goetz@6458 | 2918 | |
goetz@6458 | 2919 | // Compare char[] arrays. |
goetz@6458 | 2920 | // |
goetz@6458 | 2921 | // str1_reg USE only |
goetz@6458 | 2922 | // str2_reg USE only |
goetz@6458 | 2923 | // cnt_reg USE_DEF, due to tmp reg shortage |
goetz@6458 | 2924 | // result_reg DEF only, might compromise USE only registers |
goetz@6458 | 2925 | void MacroAssembler::char_arrays_equals(Register str1_reg, Register str2_reg, Register cnt_reg, Register result_reg, |
goetz@6458 | 2926 | Register tmp1_reg, Register tmp2_reg, Register tmp3_reg, Register tmp4_reg, |
goetz@6458 | 2927 | Register tmp5_reg) { |
goetz@6458 | 2928 | |
goetz@6458 | 2929 | // Str1 may be the same register as str2 which can occur e.g. after scalar replacement. |
goetz@6458 | 2930 | assert_different_registers(result_reg, str1_reg, cnt_reg, tmp1_reg, tmp2_reg, tmp3_reg, tmp4_reg, tmp5_reg); |
goetz@6458 | 2931 | assert_different_registers(result_reg, str2_reg, cnt_reg, tmp1_reg, tmp2_reg, tmp3_reg, tmp4_reg, tmp5_reg); |
goetz@6458 | 2932 | |
goetz@6458 | 2933 | // Offset 0 should be 32 byte aligned. |
goetz@6458 | 2934 | Label Linit_cbc, Lcbc, Lloop, Ldone_true, Ldone_false; |
goetz@6458 | 2935 | Register index_reg = tmp5_reg; |
goetz@6458 | 2936 | Register cbc_iter = tmp4_reg; |
goetz@6458 | 2937 | |
goetz@6458 | 2938 | //-1: |
goetz@6458 | 2939 | dcbtct(str1_reg, 0x00); // Indicate R/O access to str1. |
goetz@6458 | 2940 | dcbtct(str2_reg, 0x00); // Indicate R/O access to str2. |
goetz@6458 | 2941 | //1: |
goetz@6458 | 2942 | andi(cbc_iter, cnt_reg, 4-1); // Remaining iterations after 4 java characters per iteration loop. |
goetz@6458 | 2943 | li(index_reg, 0); // init |
goetz@6458 | 2944 | li(result_reg, 0); // assume false |
goetz@6458 | 2945 | srwi_(tmp2_reg, cnt_reg, exact_log2(4)); // Div: 4 java characters per iteration (main loop). |
goetz@6458 | 2946 | |
goetz@6458 | 2947 | cmpwi(CCR1, cbc_iter, 0); // CCR1 = (cbc_iter==0) |
goetz@6458 | 2948 | beq(CCR0, Linit_cbc); // too short |
goetz@6458 | 2949 | mtctr(tmp2_reg); |
goetz@6458 | 2950 | //8: |
goetz@6458 | 2951 | bind(Lloop); |
goetz@6458 | 2952 | ldx(tmp1_reg, str1_reg, index_reg); |
goetz@6458 | 2953 | ldx(tmp2_reg, str2_reg, index_reg); |
goetz@6458 | 2954 | cmpd(CCR0, tmp1_reg, tmp2_reg); |
goetz@6458 | 2955 | bne(CCR0, Ldone_false); // Unequal char pair found -> done. |
goetz@6458 | 2956 | addi(index_reg, index_reg, 4*sizeof(jchar)); |
goetz@6458 | 2957 | bdnz(Lloop); |
goetz@6458 | 2958 | //14: |
goetz@6458 | 2959 | bind(Linit_cbc); |
goetz@6458 | 2960 | beq(CCR1, Ldone_true); |
goetz@6458 | 2961 | mtctr(cbc_iter); |
goetz@6458 | 2962 | //16: |
goetz@6458 | 2963 | bind(Lcbc); |
goetz@6458 | 2964 | lhzx(tmp1_reg, str1_reg, index_reg); |
goetz@6458 | 2965 | lhzx(tmp2_reg, str2_reg, index_reg); |
goetz@6458 | 2966 | cmpw(CCR0, tmp1_reg, tmp2_reg); |
goetz@6458 | 2967 | bne(CCR0, Ldone_false); // Unequal char pair found -> done. |
goetz@6458 | 2968 | addi(index_reg, index_reg, 1*sizeof(jchar)); |
goetz@6458 | 2969 | bdnz(Lcbc); |
goetz@6458 | 2970 | nop(); |
goetz@6458 | 2971 | bind(Ldone_true); |
goetz@6458 | 2972 | li(result_reg, 1); |
goetz@6458 | 2973 | //24: |
goetz@6458 | 2974 | bind(Ldone_false); |
goetz@6458 | 2975 | } |
goetz@6458 | 2976 | |
goetz@6458 | 2977 | |
goetz@6458 | 2978 | void MacroAssembler::char_arrays_equalsImm(Register str1_reg, Register str2_reg, int cntval, Register result_reg, |
goetz@6458 | 2979 | Register tmp1_reg, Register tmp2_reg) { |
goetz@6458 | 2980 | // Str1 may be the same register as str2 which can occur e.g. after scalar replacement. |
goetz@6458 | 2981 | assert_different_registers(result_reg, str1_reg, tmp1_reg, tmp2_reg); |
goetz@6458 | 2982 | assert_different_registers(result_reg, str2_reg, tmp1_reg, tmp2_reg); |
goetz@6458 | 2983 | assert(sizeof(jchar) == 2, "must be"); |
goetz@6458 | 2984 | assert(cntval >= 0 && ((cntval & 0x7fff) == cntval), "wrong immediate"); |
goetz@6458 | 2985 | |
goetz@6458 | 2986 | Label Ldone_false; |
goetz@6458 | 2987 | |
goetz@6458 | 2988 | if (cntval < 16) { // short case |
goetz@6458 | 2989 | if (cntval != 0) li(result_reg, 0); // assume false |
goetz@6458 | 2990 | |
goetz@6458 | 2991 | const int num_bytes = cntval*sizeof(jchar); |
goetz@6458 | 2992 | int index = 0; |
goetz@6458 | 2993 | for (int next_index; (next_index = index + 8) <= num_bytes; index = next_index) { |
goetz@6458 | 2994 | ld(tmp1_reg, index, str1_reg); |
goetz@6458 | 2995 | ld(tmp2_reg, index, str2_reg); |
goetz@6458 | 2996 | cmpd(CCR0, tmp1_reg, tmp2_reg); |
goetz@6458 | 2997 | bne(CCR0, Ldone_false); |
goetz@6458 | 2998 | } |
goetz@6458 | 2999 | if (cntval & 2) { |
goetz@6458 | 3000 | lwz(tmp1_reg, index, str1_reg); |
goetz@6458 | 3001 | lwz(tmp2_reg, index, str2_reg); |
goetz@6458 | 3002 | cmpw(CCR0, tmp1_reg, tmp2_reg); |
goetz@6458 | 3003 | bne(CCR0, Ldone_false); |
goetz@6458 | 3004 | index += 4; |
goetz@6458 | 3005 | } |
goetz@6458 | 3006 | if (cntval & 1) { |
goetz@6458 | 3007 | lhz(tmp1_reg, index, str1_reg); |
goetz@6458 | 3008 | lhz(tmp2_reg, index, str2_reg); |
goetz@6458 | 3009 | cmpw(CCR0, tmp1_reg, tmp2_reg); |
goetz@6458 | 3010 | bne(CCR0, Ldone_false); |
goetz@6458 | 3011 | } |
goetz@6458 | 3012 | // fallthrough: true |
goetz@6458 | 3013 | } else { |
goetz@6458 | 3014 | Label Lloop; |
goetz@6458 | 3015 | Register index_reg = tmp1_reg; |
goetz@6458 | 3016 | const int loopcnt = cntval/4; |
goetz@6458 | 3017 | assert(loopcnt > 0, "must be"); |
goetz@6458 | 3018 | // Offset 0 should be 32 byte aligned. |
goetz@6458 | 3019 | //2: |
goetz@6458 | 3020 | dcbtct(str1_reg, 0x00); // Indicate R/O access to str1. |
goetz@6458 | 3021 | dcbtct(str2_reg, 0x00); // Indicate R/O access to str2. |
goetz@6458 | 3022 | li(tmp2_reg, loopcnt); |
goetz@6458 | 3023 | li(index_reg, 0); // init |
goetz@6458 | 3024 | li(result_reg, 0); // assume false |
goetz@6458 | 3025 | mtctr(tmp2_reg); |
goetz@6458 | 3026 | //8: |
goetz@6458 | 3027 | bind(Lloop); |
goetz@6458 | 3028 | ldx(R0, str1_reg, index_reg); |
goetz@6458 | 3029 | ldx(tmp2_reg, str2_reg, index_reg); |
goetz@6458 | 3030 | cmpd(CCR0, R0, tmp2_reg); |
goetz@6458 | 3031 | bne(CCR0, Ldone_false); // Unequal char pair found -> done. |
goetz@6458 | 3032 | addi(index_reg, index_reg, 4*sizeof(jchar)); |
goetz@6458 | 3033 | bdnz(Lloop); |
goetz@6458 | 3034 | //14: |
goetz@6458 | 3035 | if (cntval & 2) { |
goetz@6458 | 3036 | lwzx(R0, str1_reg, index_reg); |
goetz@6458 | 3037 | lwzx(tmp2_reg, str2_reg, index_reg); |
goetz@6458 | 3038 | cmpw(CCR0, R0, tmp2_reg); |
goetz@6458 | 3039 | bne(CCR0, Ldone_false); |
goetz@6458 | 3040 | if (cntval & 1) addi(index_reg, index_reg, 2*sizeof(jchar)); |
goetz@6458 | 3041 | } |
goetz@6458 | 3042 | if (cntval & 1) { |
goetz@6458 | 3043 | lhzx(R0, str1_reg, index_reg); |
goetz@6458 | 3044 | lhzx(tmp2_reg, str2_reg, index_reg); |
goetz@6458 | 3045 | cmpw(CCR0, R0, tmp2_reg); |
goetz@6458 | 3046 | bne(CCR0, Ldone_false); |
goetz@6458 | 3047 | } |
goetz@6458 | 3048 | // fallthru: true |
goetz@6458 | 3049 | } |
goetz@6458 | 3050 | li(result_reg, 1); |
goetz@6458 | 3051 | bind(Ldone_false); |
goetz@6458 | 3052 | } |
goetz@6458 | 3053 | |
gromero@9496 | 3054 | // Helpers for Intrinsic Emitters |
gromero@9496 | 3055 | // |
gromero@9496 | 3056 | // Revert the byte order of a 32bit value in a register |
gromero@9496 | 3057 | // src: 0x44556677 |
gromero@9496 | 3058 | // dst: 0x77665544 |
gromero@9496 | 3059 | // Three steps to obtain the result: |
gromero@9496 | 3060 | // 1) Rotate src (as doubleword) left 5 bytes. That puts the leftmost byte of the src word |
gromero@9496 | 3061 | // into the rightmost byte position. Afterwards, everything left of the rightmost byte is cleared. |
gromero@9496 | 3062 | // This value initializes dst. |
gromero@9496 | 3063 | // 2) Rotate src (as word) left 3 bytes. That puts the rightmost byte of the src word into the leftmost |
gromero@9496 | 3064 | // byte position. Furthermore, byte 5 is rotated into byte 6 position where it is supposed to go. |
gromero@9496 | 3065 | // This value is mask inserted into dst with a [0..23] mask of 1s. |
gromero@9496 | 3066 | // 3) Rotate src (as word) left 1 byte. That puts byte 6 into byte 5 position. |
gromero@9496 | 3067 | // This value is mask inserted into dst with a [8..15] mask of 1s. |
gromero@9496 | 3068 | void MacroAssembler::load_reverse_32(Register dst, Register src) { |
gromero@9496 | 3069 | assert_different_registers(dst, src); |
gromero@9496 | 3070 | |
gromero@9496 | 3071 | rldicl(dst, src, (4+1)*8, 56); // Rotate byte 4 into position 7 (rightmost), clear all to the left. |
gromero@9496 | 3072 | rlwimi(dst, src, 3*8, 0, 23); // Insert byte 5 into position 6, 7 into 4, leave pos 7 alone. |
gromero@9496 | 3073 | rlwimi(dst, src, 1*8, 8, 15); // Insert byte 6 into position 5, leave the rest alone. |
gromero@9496 | 3074 | } |
gromero@9496 | 3075 | |
gromero@9496 | 3076 | // Calculate the column addresses of the crc32 lookup table into distinct registers. |
gromero@9496 | 3077 | // This loop-invariant calculation is moved out of the loop body, reducing the loop |
gromero@9496 | 3078 | // body size from 20 to 16 instructions. |
gromero@9496 | 3079 | // Returns the offset that was used to calculate the address of column tc3. |
gromero@9496 | 3080 | // Due to register shortage, setting tc3 may overwrite table. With the return offset |
gromero@9496 | 3081 | // at hand, the original table address can be easily reconstructed. |
gromero@9496 | 3082 | int MacroAssembler::crc32_table_columns(Register table, Register tc0, Register tc1, Register tc2, Register tc3) { |
gromero@9496 | 3083 | |
gromero@9496 | 3084 | #ifdef VM_LITTLE_ENDIAN |
gromero@9496 | 3085 | // This is what we implement (the DOLIT4 part): |
gromero@9496 | 3086 | // ========================================================================= */ |
gromero@9496 | 3087 | // #define DOLIT4 c ^= *buf4++; \ |
gromero@9496 | 3088 | // c = crc_table[3][c & 0xff] ^ crc_table[2][(c >> 8) & 0xff] ^ \ |
gromero@9496 | 3089 | // crc_table[1][(c >> 16) & 0xff] ^ crc_table[0][c >> 24] |
gromero@9496 | 3090 | // #define DOLIT32 DOLIT4; DOLIT4; DOLIT4; DOLIT4; DOLIT4; DOLIT4; DOLIT4; DOLIT4 |
gromero@9496 | 3091 | // ========================================================================= */ |
gromero@9496 | 3092 | const int ix0 = 3*(4*CRC32_COLUMN_SIZE); |
gromero@9496 | 3093 | const int ix1 = 2*(4*CRC32_COLUMN_SIZE); |
gromero@9496 | 3094 | const int ix2 = 1*(4*CRC32_COLUMN_SIZE); |
gromero@9496 | 3095 | const int ix3 = 0*(4*CRC32_COLUMN_SIZE); |
gromero@9496 | 3096 | #else |
gromero@9496 | 3097 | // This is what we implement (the DOBIG4 part): |
gromero@9496 | 3098 | // ========================================================================= |
gromero@9496 | 3099 | // #define DOBIG4 c ^= *++buf4; \ |
gromero@9496 | 3100 | // c = crc_table[4][c & 0xff] ^ crc_table[5][(c >> 8) & 0xff] ^ \ |
gromero@9496 | 3101 | // crc_table[6][(c >> 16) & 0xff] ^ crc_table[7][c >> 24] |
gromero@9496 | 3102 | // #define DOBIG32 DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4 |
gromero@9496 | 3103 | // ========================================================================= |
gromero@9496 | 3104 | const int ix0 = 4*(4*CRC32_COLUMN_SIZE); |
gromero@9496 | 3105 | const int ix1 = 5*(4*CRC32_COLUMN_SIZE); |
gromero@9496 | 3106 | const int ix2 = 6*(4*CRC32_COLUMN_SIZE); |
gromero@9496 | 3107 | const int ix3 = 7*(4*CRC32_COLUMN_SIZE); |
gromero@9496 | 3108 | #endif |
gromero@9496 | 3109 | assert_different_registers(table, tc0, tc1, tc2); |
gromero@9496 | 3110 | assert(table == tc3, "must be!"); |
gromero@9496 | 3111 | |
gromero@9496 | 3112 | if (ix0 != 0) addi(tc0, table, ix0); |
gromero@9496 | 3113 | if (ix1 != 0) addi(tc1, table, ix1); |
gromero@9496 | 3114 | if (ix2 != 0) addi(tc2, table, ix2); |
gromero@9496 | 3115 | if (ix3 != 0) addi(tc3, table, ix3); |
gromero@9496 | 3116 | |
gromero@9496 | 3117 | return ix3; |
gromero@9496 | 3118 | } |
gromero@9496 | 3119 | |
gromero@9496 | 3120 | /** |
gromero@9496 | 3121 | * uint32_t crc; |
gromero@9496 | 3122 | * timesXtoThe32[crc & 0xFF] ^ (crc >> 8); |
gromero@9496 | 3123 | */ |
gromero@9496 | 3124 | void MacroAssembler::fold_byte_crc32(Register crc, Register val, Register table, Register tmp) { |
gromero@9496 | 3125 | assert_different_registers(crc, table, tmp); |
gromero@9496 | 3126 | assert_different_registers(val, table); |
gromero@9496 | 3127 | |
gromero@9496 | 3128 | if (crc == val) { // Must rotate first to use the unmodified value. |
gromero@9496 | 3129 | rlwinm(tmp, val, 2, 24-2, 31-2); // Insert (rightmost) byte 7 of val, shifted left by 2, into byte 6..7 of tmp, clear the rest. |
gromero@9496 | 3130 | // As we use a word (4-byte) instruction, we have to adapt the mask bit positions. |
gromero@9496 | 3131 | srwi(crc, crc, 8); // Unsigned shift, clear leftmost 8 bits. |
gromero@9496 | 3132 | } else { |
gromero@9496 | 3133 | srwi(crc, crc, 8); // Unsigned shift, clear leftmost 8 bits. |
gromero@9496 | 3134 | rlwinm(tmp, val, 2, 24-2, 31-2); // Insert (rightmost) byte 7 of val, shifted left by 2, into byte 6..7 of tmp, clear the rest. |
gromero@9496 | 3135 | } |
gromero@9496 | 3136 | lwzx(tmp, table, tmp); |
gromero@9496 | 3137 | xorr(crc, crc, tmp); |
gromero@9496 | 3138 | } |
gromero@9496 | 3139 | |
gromero@9496 | 3140 | /** |
gromero@9496 | 3141 | * uint32_t crc; |
gromero@9496 | 3142 | * timesXtoThe32[crc & 0xFF] ^ (crc >> 8); |
gromero@9496 | 3143 | */ |
gromero@9496 | 3144 | void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) { |
gromero@9496 | 3145 | fold_byte_crc32(crc, crc, table, tmp); |
gromero@9496 | 3146 | } |
gromero@9496 | 3147 | |
gromero@9496 | 3148 | /** |
gromero@9496 | 3149 | * Emits code to update CRC-32 with a byte value according to constants in table. |
gromero@9496 | 3150 | * |
gromero@9496 | 3151 | * @param [in,out]crc Register containing the crc. |
gromero@9496 | 3152 | * @param [in]val Register containing the byte to fold into the CRC. |
gromero@9496 | 3153 | * @param [in]table Register containing the table of crc constants. |
gromero@9496 | 3154 | * |
gromero@9496 | 3155 | * uint32_t crc; |
gromero@9496 | 3156 | * val = crc_table[(val ^ crc) & 0xFF]; |
gromero@9496 | 3157 | * crc = val ^ (crc >> 8); |
gromero@9496 | 3158 | */ |
gromero@9496 | 3159 | void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) { |
gromero@9496 | 3160 | BLOCK_COMMENT("update_byte_crc32:"); |
gromero@9496 | 3161 | xorr(val, val, crc); |
gromero@9496 | 3162 | fold_byte_crc32(crc, val, table, val); |
gromero@9496 | 3163 | } |
gromero@9496 | 3164 | |
gromero@9496 | 3165 | /** |
gromero@9496 | 3166 | * @param crc register containing existing CRC (32-bit) |
gromero@9496 | 3167 | * @param buf register pointing to input byte buffer (byte*) |
gromero@9496 | 3168 | * @param len register containing number of bytes |
gromero@9496 | 3169 | * @param table register pointing to CRC table |
gromero@9496 | 3170 | */ |
gromero@9496 | 3171 | void MacroAssembler::update_byteLoop_crc32(Register crc, Register buf, Register len, Register table, |
gromero@9496 | 3172 | Register data, bool loopAlignment, bool invertCRC) { |
gromero@9496 | 3173 | assert_different_registers(crc, buf, len, table, data); |
gromero@9496 | 3174 | |
gromero@9496 | 3175 | Label L_mainLoop, L_done; |
gromero@9496 | 3176 | const int mainLoop_stepping = 1; |
gromero@9496 | 3177 | const int mainLoop_alignment = loopAlignment ? 32 : 4; // (InputForNewCode > 4 ? InputForNewCode : 32) : 4; |
gromero@9496 | 3178 | |
gromero@9496 | 3179 | // Process all bytes in a single-byte loop. |
gromero@9496 | 3180 | cmpdi(CCR0, len, 0); // Anything to do? |
gromero@9496 | 3181 | mtctr(len); |
gromero@9496 | 3182 | beq(CCR0, L_done); |
gromero@9496 | 3183 | |
gromero@9496 | 3184 | if (invertCRC) { |
gromero@9496 | 3185 | nand(crc, crc, crc); // ~c |
gromero@9496 | 3186 | } |
gromero@9496 | 3187 | |
gromero@9496 | 3188 | align(mainLoop_alignment); |
gromero@9496 | 3189 | BIND(L_mainLoop); |
gromero@9496 | 3190 | lbz(data, 0, buf); // Byte from buffer, zero-extended. |
gromero@9496 | 3191 | addi(buf, buf, mainLoop_stepping); // Advance buffer position. |
gromero@9496 | 3192 | update_byte_crc32(crc, data, table); |
gromero@9496 | 3193 | bdnz(L_mainLoop); // Iterate. |
gromero@9496 | 3194 | |
gromero@9496 | 3195 | if (invertCRC) { |
gromero@9496 | 3196 | nand(crc, crc, crc); // ~c |
gromero@9496 | 3197 | } |
gromero@9496 | 3198 | |
gromero@9496 | 3199 | bind(L_done); |
gromero@9496 | 3200 | } |
gromero@9496 | 3201 | |
gromero@9496 | 3202 | /** |
gromero@9496 | 3203 | * Emits code to update CRC-32 with a 4-byte value according to constants in table |
gromero@9496 | 3204 | * Implementation according to jdk/src/share/native/java/util/zip/zlib-1.2.8/crc32.c |
gromero@9496 | 3205 | */ |
gromero@9496 | 3206 | // A not on the lookup table address(es): |
gromero@9496 | 3207 | // The lookup table consists of two sets of four columns each. |
gromero@9496 | 3208 | // The columns {0..3} are used for little-endian machines. |
gromero@9496 | 3209 | // The columns {4..7} are used for big-endian machines. |
gromero@9496 | 3210 | // To save the effort of adding the column offset to the table address each time |
gromero@9496 | 3211 | // a table element is looked up, it is possible to pass the pre-calculated |
gromero@9496 | 3212 | // column addresses. |
gromero@9496 | 3213 | // Uses R9..R12 as work register. Must be saved/restored by caller, if necessary. |
gromero@9496 | 3214 | void MacroAssembler::update_1word_crc32(Register crc, Register buf, Register table, int bufDisp, int bufInc, |
gromero@9496 | 3215 | Register t0, Register t1, Register t2, Register t3, |
gromero@9496 | 3216 | Register tc0, Register tc1, Register tc2, Register tc3) { |
gromero@9496 | 3217 | assert_different_registers(crc, t3); |
gromero@9496 | 3218 | |
gromero@9496 | 3219 | // XOR crc with next four bytes of buffer. |
gromero@9496 | 3220 | lwz(t3, bufDisp, buf); |
gromero@9496 | 3221 | if (bufInc != 0) { |
gromero@9496 | 3222 | addi(buf, buf, bufInc); |
gromero@9496 | 3223 | } |
gromero@9496 | 3224 | xorr(t3, t3, crc); |
gromero@9496 | 3225 | |
gromero@9496 | 3226 | // Chop crc into 4 single-byte pieces, shifted left 2 bits, to form the table indices. |
gromero@9496 | 3227 | rlwinm(t0, t3, 2, 24-2, 31-2); // ((t1 >> 0) & 0xff) << 2 |
gromero@9496 | 3228 | rlwinm(t1, t3, 32+(2- 8), 24-2, 31-2); // ((t1 >> 8) & 0xff) << 2 |
gromero@9496 | 3229 | rlwinm(t2, t3, 32+(2-16), 24-2, 31-2); // ((t1 >> 16) & 0xff) << 2 |
gromero@9496 | 3230 | rlwinm(t3, t3, 32+(2-24), 24-2, 31-2); // ((t1 >> 24) & 0xff) << 2 |
gromero@9496 | 3231 | |
gromero@9496 | 3232 | // Use the pre-calculated column addresses. |
gromero@9496 | 3233 | // Load pre-calculated table values. |
gromero@9496 | 3234 | lwzx(t0, tc0, t0); |
gromero@9496 | 3235 | lwzx(t1, tc1, t1); |
gromero@9496 | 3236 | lwzx(t2, tc2, t2); |
gromero@9496 | 3237 | lwzx(t3, tc3, t3); |
gromero@9496 | 3238 | |
gromero@9496 | 3239 | // Calculate new crc from table values. |
gromero@9496 | 3240 | xorr(t0, t0, t1); |
gromero@9496 | 3241 | xorr(t2, t2, t3); |
gromero@9496 | 3242 | xorr(crc, t0, t2); // Now crc contains the final checksum value. |
gromero@9496 | 3243 | } |
gromero@9496 | 3244 | |
gromero@9496 | 3245 | /** |
gromero@9496 | 3246 | * @param crc register containing existing CRC (32-bit) |
gromero@9496 | 3247 | * @param buf register pointing to input byte buffer (byte*) |
gromero@9496 | 3248 | * @param len register containing number of bytes |
gromero@9496 | 3249 | * @param table register pointing to CRC table |
gromero@9496 | 3250 | * |
gromero@9496 | 3251 | * Uses R9..R12 as work register. Must be saved/restored by caller! |
gromero@9496 | 3252 | */ |
gromero@9496 | 3253 | void MacroAssembler::kernel_crc32_2word(Register crc, Register buf, Register len, Register table, |
gromero@9496 | 3254 | Register t0, Register t1, Register t2, Register t3, |
gromero@9496 | 3255 | Register tc0, Register tc1, Register tc2, Register tc3) { |
gromero@9496 | 3256 | assert_different_registers(crc, buf, len, table); |
gromero@9496 | 3257 | |
gromero@9496 | 3258 | Label L_mainLoop, L_tail; |
gromero@9496 | 3259 | Register tmp = t0; |
gromero@9496 | 3260 | Register data = t0; |
gromero@9496 | 3261 | Register tmp2 = t1; |
gromero@9496 | 3262 | const int mainLoop_stepping = 8; |
gromero@9496 | 3263 | const int tailLoop_stepping = 1; |
gromero@9496 | 3264 | const int log_stepping = exact_log2(mainLoop_stepping); |
gromero@9496 | 3265 | const int mainLoop_alignment = 32; // InputForNewCode > 4 ? InputForNewCode : 32; |
gromero@9496 | 3266 | const int complexThreshold = 2*mainLoop_stepping; |
gromero@9496 | 3267 | |
gromero@9496 | 3268 | // Don't test for len <= 0 here. This pathological case should not occur anyway. |
gromero@9496 | 3269 | // Optimizing for it by adding a test and a branch seems to be a waste of CPU cycles. |
gromero@9496 | 3270 | // The situation itself is detected and handled correctly by the conditional branches |
gromero@9496 | 3271 | // following aghi(len, -stepping) and aghi(len, +stepping). |
gromero@9496 | 3272 | assert(tailLoop_stepping == 1, "check tailLoop_stepping!"); |
gromero@9496 | 3273 | |
gromero@9496 | 3274 | BLOCK_COMMENT("kernel_crc32_2word {"); |
gromero@9496 | 3275 | |
gromero@9496 | 3276 | nand(crc, crc, crc); // ~c |
gromero@9496 | 3277 | |
gromero@9496 | 3278 | // Check for short (<mainLoop_stepping) buffer. |
gromero@9496 | 3279 | cmpdi(CCR0, len, complexThreshold); |
gromero@9496 | 3280 | blt(CCR0, L_tail); |
gromero@9496 | 3281 | |
gromero@9496 | 3282 | // Pre-mainLoop alignment did show a slight (1%) positive effect on performance. |
gromero@9496 | 3283 | // We leave the code in for reference. Maybe we need alignment when we exploit vector instructions. |
gromero@9496 | 3284 | { |
gromero@9496 | 3285 | // Align buf addr to mainLoop_stepping boundary. |
gromero@9496 | 3286 | neg(tmp2, buf); // Calculate # preLoop iterations for alignment. |
gromero@9496 | 3287 | rldicl(tmp2, tmp2, 0, 64-log_stepping); // Rotate tmp2 0 bits, insert into tmp2, anding with mask with 1s from 62..63. |
gromero@9496 | 3288 | |
gromero@9496 | 3289 | if (complexThreshold > mainLoop_stepping) { |
gromero@9496 | 3290 | sub(len, len, tmp2); // Remaining bytes for main loop (>=mainLoop_stepping is guaranteed). |
gromero@9496 | 3291 | } else { |
gromero@9496 | 3292 | sub(tmp, len, tmp2); // Remaining bytes for main loop. |
gromero@9496 | 3293 | cmpdi(CCR0, tmp, mainLoop_stepping); |
gromero@9496 | 3294 | blt(CCR0, L_tail); // For less than one mainloop_stepping left, do only tail processing |
gromero@9496 | 3295 | mr(len, tmp); // remaining bytes for main loop (>=mainLoop_stepping is guaranteed). |
gromero@9496 | 3296 | } |
gromero@9496 | 3297 | update_byteLoop_crc32(crc, buf, tmp2, table, data, false, false); |
gromero@9496 | 3298 | } |
gromero@9496 | 3299 | |
gromero@9496 | 3300 | srdi(tmp2, len, log_stepping); // #iterations for mainLoop |
gromero@9496 | 3301 | andi(len, len, mainLoop_stepping-1); // remaining bytes for tailLoop |
gromero@9496 | 3302 | mtctr(tmp2); |
gromero@9496 | 3303 | |
gromero@9496 | 3304 | #ifdef VM_LITTLE_ENDIAN |
gromero@9496 | 3305 | Register crc_rv = crc; |
gromero@9496 | 3306 | #else |
gromero@9496 | 3307 | Register crc_rv = tmp; // Load_reverse needs separate registers to work on. |
gromero@9496 | 3308 | // Occupies tmp, but frees up crc. |
gromero@9496 | 3309 | load_reverse_32(crc_rv, crc); // Revert byte order because we are dealing with big-endian data. |
gromero@9496 | 3310 | tmp = crc; |
gromero@9496 | 3311 | #endif |
gromero@9496 | 3312 | |
gromero@9496 | 3313 | int reconstructTableOffset = crc32_table_columns(table, tc0, tc1, tc2, tc3); |
gromero@9496 | 3314 | |
gromero@9496 | 3315 | align(mainLoop_alignment); // Octoword-aligned loop address. Shows 2% improvement. |
gromero@9496 | 3316 | BIND(L_mainLoop); |
gromero@9496 | 3317 | update_1word_crc32(crc_rv, buf, table, 0, 0, crc_rv, t1, t2, t3, tc0, tc1, tc2, tc3); |
gromero@9496 | 3318 | update_1word_crc32(crc_rv, buf, table, 4, mainLoop_stepping, crc_rv, t1, t2, t3, tc0, tc1, tc2, tc3); |
gromero@9496 | 3319 | bdnz(L_mainLoop); |
gromero@9496 | 3320 | |
gromero@9496 | 3321 | #ifndef VM_LITTLE_ENDIAN |
gromero@9496 | 3322 | load_reverse_32(crc, crc_rv); // Revert byte order because we are dealing with big-endian data. |
gromero@9496 | 3323 | tmp = crc_rv; // Tmp uses it's original register again. |
gromero@9496 | 3324 | #endif |
gromero@9496 | 3325 | |
gromero@9496 | 3326 | // Restore original table address for tailLoop. |
gromero@9496 | 3327 | if (reconstructTableOffset != 0) { |
gromero@9496 | 3328 | addi(table, table, -reconstructTableOffset); |
gromero@9496 | 3329 | } |
gromero@9496 | 3330 | |
gromero@9496 | 3331 | // Process last few (<complexThreshold) bytes of buffer. |
gromero@9496 | 3332 | BIND(L_tail); |
gromero@9496 | 3333 | update_byteLoop_crc32(crc, buf, len, table, data, false, false); |
gromero@9496 | 3334 | |
gromero@9496 | 3335 | nand(crc, crc, crc); // ~c |
gromero@9496 | 3336 | BLOCK_COMMENT("} kernel_crc32_2word"); |
gromero@9496 | 3337 | } |
gromero@9496 | 3338 | |
gromero@9496 | 3339 | /** |
gromero@9496 | 3340 | * @param crc register containing existing CRC (32-bit) |
gromero@9496 | 3341 | * @param buf register pointing to input byte buffer (byte*) |
gromero@9496 | 3342 | * @param len register containing number of bytes |
gromero@9496 | 3343 | * @param table register pointing to CRC table |
gromero@9496 | 3344 | * |
gromero@9496 | 3345 | * uses R9..R12 as work register. Must be saved/restored by caller! |
gromero@9496 | 3346 | */ |
gromero@9496 | 3347 | void MacroAssembler::kernel_crc32_1word(Register crc, Register buf, Register len, Register table, |
gromero@9496 | 3348 | Register t0, Register t1, Register t2, Register t3, |
gromero@9496 | 3349 | Register tc0, Register tc1, Register tc2, Register tc3) { |
gromero@9496 | 3350 | assert_different_registers(crc, buf, len, table); |
gromero@9496 | 3351 | |
gromero@9496 | 3352 | Label L_mainLoop, L_tail; |
gromero@9496 | 3353 | Register tmp = t0; |
gromero@9496 | 3354 | Register data = t0; |
gromero@9496 | 3355 | Register tmp2 = t1; |
gromero@9496 | 3356 | const int mainLoop_stepping = 4; |
gromero@9496 | 3357 | const int tailLoop_stepping = 1; |
gromero@9496 | 3358 | const int log_stepping = exact_log2(mainLoop_stepping); |
gromero@9496 | 3359 | const int mainLoop_alignment = 32; // InputForNewCode > 4 ? InputForNewCode : 32; |
gromero@9496 | 3360 | const int complexThreshold = 2*mainLoop_stepping; |
gromero@9496 | 3361 | |
gromero@9496 | 3362 | // Don't test for len <= 0 here. This pathological case should not occur anyway. |
gromero@9496 | 3363 | // Optimizing for it by adding a test and a branch seems to be a waste of CPU cycles. |
gromero@9496 | 3364 | // The situation itself is detected and handled correctly by the conditional branches |
gromero@9496 | 3365 | // following aghi(len, -stepping) and aghi(len, +stepping). |
gromero@9496 | 3366 | assert(tailLoop_stepping == 1, "check tailLoop_stepping!"); |
gromero@9496 | 3367 | |
gromero@9496 | 3368 | BLOCK_COMMENT("kernel_crc32_1word {"); |
gromero@9496 | 3369 | |
gromero@9496 | 3370 | nand(crc, crc, crc); // ~c |
gromero@9496 | 3371 | |
gromero@9496 | 3372 | // Check for short (<mainLoop_stepping) buffer. |
gromero@9496 | 3373 | cmpdi(CCR0, len, complexThreshold); |
gromero@9496 | 3374 | blt(CCR0, L_tail); |
gromero@9496 | 3375 | |
gromero@9496 | 3376 | // Pre-mainLoop alignment did show a slight (1%) positive effect on performance. |
gromero@9496 | 3377 | // We leave the code in for reference. Maybe we need alignment when we exploit vector instructions. |
gromero@9496 | 3378 | { |
gromero@9496 | 3379 | // Align buf addr to mainLoop_stepping boundary. |
gromero@9496 | 3380 | neg(tmp2, buf); // Calculate # preLoop iterations for alignment. |
gromero@9496 | 3381 | rldicl(tmp2, tmp2, 0, 64-log_stepping); // Rotate tmp2 0 bits, insert into tmp2, anding with mask with 1s from 62..63. |
gromero@9496 | 3382 | |
gromero@9496 | 3383 | if (complexThreshold > mainLoop_stepping) { |
gromero@9496 | 3384 | sub(len, len, tmp2); // Remaining bytes for main loop (>=mainLoop_stepping is guaranteed). |
gromero@9496 | 3385 | } else { |
gromero@9496 | 3386 | sub(tmp, len, tmp2); // Remaining bytes for main loop. |
gromero@9496 | 3387 | cmpdi(CCR0, tmp, mainLoop_stepping); |
gromero@9496 | 3388 | blt(CCR0, L_tail); // For less than one mainloop_stepping left, do only tail processing |
gromero@9496 | 3389 | mr(len, tmp); // remaining bytes for main loop (>=mainLoop_stepping is guaranteed). |
gromero@9496 | 3390 | } |
gromero@9496 | 3391 | update_byteLoop_crc32(crc, buf, tmp2, table, data, false, false); |
gromero@9496 | 3392 | } |
gromero@9496 | 3393 | |
gromero@9496 | 3394 | srdi(tmp2, len, log_stepping); // #iterations for mainLoop |
gromero@9496 | 3395 | andi(len, len, mainLoop_stepping-1); // remaining bytes for tailLoop |
gromero@9496 | 3396 | mtctr(tmp2); |
gromero@9496 | 3397 | |
gromero@9496 | 3398 | #ifdef VM_LITTLE_ENDIAN |
gromero@9496 | 3399 | Register crc_rv = crc; |
gromero@9496 | 3400 | #else |
gromero@9496 | 3401 | Register crc_rv = tmp; // Load_reverse needs separate registers to work on. |
gromero@9496 | 3402 | // Occupies tmp, but frees up crc. |
gromero@9496 | 3403 | load_reverse_32(crc_rv, crc); // evert byte order because we are dealing with big-endian data. |
gromero@9496 | 3404 | tmp = crc; |
gromero@9496 | 3405 | #endif |
gromero@9496 | 3406 | |
gromero@9496 | 3407 | int reconstructTableOffset = crc32_table_columns(table, tc0, tc1, tc2, tc3); |
gromero@9496 | 3408 | |
gromero@9496 | 3409 | align(mainLoop_alignment); // Octoword-aligned loop address. Shows 2% improvement. |
gromero@9496 | 3410 | BIND(L_mainLoop); |
gromero@9496 | 3411 | update_1word_crc32(crc_rv, buf, table, 0, mainLoop_stepping, crc_rv, t1, t2, t3, tc0, tc1, tc2, tc3); |
gromero@9496 | 3412 | bdnz(L_mainLoop); |
gromero@9496 | 3413 | |
gromero@9496 | 3414 | #ifndef VM_LITTLE_ENDIAN |
gromero@9496 | 3415 | load_reverse_32(crc, crc_rv); // Revert byte order because we are dealing with big-endian data. |
gromero@9496 | 3416 | tmp = crc_rv; // Tmp uses it's original register again. |
gromero@9496 | 3417 | #endif |
gromero@9496 | 3418 | |
gromero@9496 | 3419 | // Restore original table address for tailLoop. |
gromero@9496 | 3420 | if (reconstructTableOffset != 0) { |
gromero@9496 | 3421 | addi(table, table, -reconstructTableOffset); |
gromero@9496 | 3422 | } |
gromero@9496 | 3423 | |
gromero@9496 | 3424 | // Process last few (<complexThreshold) bytes of buffer. |
gromero@9496 | 3425 | BIND(L_tail); |
gromero@9496 | 3426 | update_byteLoop_crc32(crc, buf, len, table, data, false, false); |
gromero@9496 | 3427 | |
gromero@9496 | 3428 | nand(crc, crc, crc); // ~c |
gromero@9496 | 3429 | BLOCK_COMMENT("} kernel_crc32_1word"); |
gromero@9496 | 3430 | } |
gromero@9496 | 3431 | |
gromero@9496 | 3432 | /** |
gromero@9496 | 3433 | * @param crc register containing existing CRC (32-bit) |
gromero@9496 | 3434 | * @param buf register pointing to input byte buffer (byte*) |
gromero@9496 | 3435 | * @param len register containing number of bytes |
gromero@9496 | 3436 | * @param table register pointing to CRC table |
gromero@9496 | 3437 | * |
gromero@9496 | 3438 | * Uses R7_ARG5, R8_ARG6 as work registers. |
gromero@9496 | 3439 | */ |
gromero@9496 | 3440 | void MacroAssembler::kernel_crc32_1byte(Register crc, Register buf, Register len, Register table, |
gromero@9496 | 3441 | Register t0, Register t1, Register t2, Register t3) { |
gromero@9496 | 3442 | assert_different_registers(crc, buf, len, table); |
gromero@9496 | 3443 | |
gromero@9496 | 3444 | Register data = t0; // Holds the current byte to be folded into crc. |
gromero@9496 | 3445 | |
gromero@9496 | 3446 | BLOCK_COMMENT("kernel_crc32_1byte {"); |
gromero@9496 | 3447 | |
gromero@9496 | 3448 | // Process all bytes in a single-byte loop. |
gromero@9496 | 3449 | update_byteLoop_crc32(crc, buf, len, table, data, true, true); |
gromero@9496 | 3450 | |
gromero@9496 | 3451 | BLOCK_COMMENT("} kernel_crc32_1byte"); |
gromero@9496 | 3452 | } |
gromero@9496 | 3453 | |
mdoerr@9497 | 3454 | /** |
mdoerr@9497 | 3455 | * @param crc register containing existing CRC (32-bit) |
mdoerr@9497 | 3456 | * @param buf register pointing to input byte buffer (byte*) |
mdoerr@9497 | 3457 | * @param len register containing number of bytes |
mdoerr@9497 | 3458 | * @param table register pointing to CRC table |
mdoerr@9497 | 3459 | * @param constants register pointing to CRC table for 128-bit aligned memory |
mdoerr@9497 | 3460 | * @param barretConstants register pointing to table for barrett reduction |
mdoerr@9497 | 3461 | * @param t0 volatile register |
mdoerr@9497 | 3462 | * @param t1 volatile register |
mdoerr@9497 | 3463 | * @param t2 volatile register |
mdoerr@9497 | 3464 | * @param t3 volatile register |
mdoerr@9497 | 3465 | */ |
mdoerr@9497 | 3466 | void MacroAssembler::kernel_crc32_1word_vpmsumd(Register crc, Register buf, Register len, Register table, |
mdoerr@9497 | 3467 | Register constants, Register barretConstants, |
mdoerr@9497 | 3468 | Register t0, Register t1, Register t2, Register t3, Register t4) { |
mdoerr@9497 | 3469 | assert_different_registers(crc, buf, len, table); |
mdoerr@9497 | 3470 | |
mdoerr@9497 | 3471 | Label L_alignedHead, L_tail, L_alignTail, L_start, L_end; |
mdoerr@9497 | 3472 | |
mdoerr@9497 | 3473 | Register prealign = t0; |
mdoerr@9497 | 3474 | Register postalign = t0; |
mdoerr@9497 | 3475 | |
mdoerr@9497 | 3476 | BLOCK_COMMENT("kernel_crc32_1word_vpmsumb {"); |
mdoerr@9497 | 3477 | |
mdoerr@9497 | 3478 | // 1. use kernel_crc32_1word for shorter than 384bit |
mdoerr@9497 | 3479 | clrldi(len, len, 32); |
mdoerr@9497 | 3480 | cmpdi(CCR0, len, 384); |
mdoerr@9497 | 3481 | bge(CCR0, L_start); |
mdoerr@9497 | 3482 | |
mdoerr@9497 | 3483 | Register tc0 = t4; |
mdoerr@9497 | 3484 | Register tc1 = constants; |
mdoerr@9497 | 3485 | Register tc2 = barretConstants; |
mdoerr@9497 | 3486 | kernel_crc32_1word(crc, buf, len, table,t0, t1, t2, t3, tc0, tc1, tc2, table); |
mdoerr@9497 | 3487 | b(L_end); |
mdoerr@9497 | 3488 | |
mdoerr@9497 | 3489 | BIND(L_start); |
mdoerr@9497 | 3490 | |
mdoerr@9497 | 3491 | // 2. ~c |
mdoerr@9497 | 3492 | nand(crc, crc, crc); |
mdoerr@9497 | 3493 | |
mdoerr@9497 | 3494 | // 3. calculate from 0 to first 128bit-aligned address |
mdoerr@9497 | 3495 | clrldi_(prealign, buf, 57); |
mdoerr@9497 | 3496 | beq(CCR0, L_alignedHead); |
mdoerr@9497 | 3497 | |
mdoerr@9497 | 3498 | subfic(prealign, prealign, 128); |
mdoerr@9497 | 3499 | |
mdoerr@9497 | 3500 | subf(len, prealign, len); |
mdoerr@9497 | 3501 | update_byteLoop_crc32(crc, buf, prealign, table, t2, false, false); |
mdoerr@9497 | 3502 | |
mdoerr@9497 | 3503 | // 4. calculate from first 128bit-aligned address to last 128bit-aligned address |
mdoerr@9497 | 3504 | BIND(L_alignedHead); |
mdoerr@9497 | 3505 | |
mdoerr@9497 | 3506 | clrldi(postalign, len, 57); |
mdoerr@9497 | 3507 | subf(len, postalign, len); |
mdoerr@9497 | 3508 | |
mdoerr@9497 | 3509 | // len must be more than 256bit |
mdoerr@9497 | 3510 | kernel_crc32_1word_aligned(crc, buf, len, constants, barretConstants, t1, t2, t3); |
mdoerr@9497 | 3511 | |
mdoerr@9497 | 3512 | // 5. calculate remaining |
mdoerr@9497 | 3513 | cmpdi(CCR0, postalign, 0); |
mdoerr@9497 | 3514 | beq(CCR0, L_tail); |
mdoerr@9497 | 3515 | |
mdoerr@9497 | 3516 | update_byteLoop_crc32(crc, buf, postalign, table, t2, false, false); |
mdoerr@9497 | 3517 | |
mdoerr@9497 | 3518 | BIND(L_tail); |
mdoerr@9497 | 3519 | |
mdoerr@9497 | 3520 | // 6. ~c |
mdoerr@9497 | 3521 | nand(crc, crc, crc); |
mdoerr@9497 | 3522 | |
mdoerr@9497 | 3523 | BIND(L_end); |
mdoerr@9497 | 3524 | |
mdoerr@9497 | 3525 | BLOCK_COMMENT("} kernel_crc32_1word_vpmsumb"); |
mdoerr@9497 | 3526 | } |
mdoerr@9497 | 3527 | |
mdoerr@9497 | 3528 | /** |
mdoerr@9497 | 3529 | * @param crc register containing existing CRC (32-bit) |
mdoerr@9497 | 3530 | * @param buf register pointing to input byte buffer (byte*) |
mdoerr@9497 | 3531 | * @param len register containing number of bytes |
mdoerr@9497 | 3532 | * @param constants register pointing to CRC table for 128-bit aligned memory |
mdoerr@9497 | 3533 | * @param barretConstants register pointing to table for barrett reduction |
mdoerr@9497 | 3534 | * @param t0 volatile register |
mdoerr@9497 | 3535 | * @param t1 volatile register |
mdoerr@9497 | 3536 | * @param t2 volatile register |
mdoerr@9497 | 3537 | */ |
mdoerr@9497 | 3538 | void MacroAssembler::kernel_crc32_1word_aligned(Register crc, Register buf, Register len, |
mdoerr@9497 | 3539 | Register constants, Register barretConstants, Register t0, Register t1, Register t2) { |
mdoerr@9497 | 3540 | Label L_mainLoop, L_tail, L_alignTail, L_barrett_reduction, L_end, L_first_warm_up_done, L_first_cool_down, L_second_cool_down, L_XOR, L_test; |
mdoerr@9497 | 3541 | Label L_lv0, L_lv1, L_lv2, L_lv3, L_lv4, L_lv5, L_lv6, L_lv7, L_lv8, L_lv9, L_lv10, L_lv11, L_lv12, L_lv13, L_lv14, L_lv15; |
mdoerr@9497 | 3542 | Label L_1, L_2, L_3, L_4; |
mdoerr@9497 | 3543 | |
mdoerr@9497 | 3544 | Register rLoaded = t0; |
mdoerr@9497 | 3545 | Register rTmp1 = t1; |
mdoerr@9497 | 3546 | Register rTmp2 = t2; |
mdoerr@9497 | 3547 | Register off16 = R22; |
mdoerr@9497 | 3548 | Register off32 = R23; |
mdoerr@9497 | 3549 | Register off48 = R24; |
mdoerr@9497 | 3550 | Register off64 = R25; |
mdoerr@9497 | 3551 | Register off80 = R26; |
mdoerr@9497 | 3552 | Register off96 = R27; |
mdoerr@9497 | 3553 | Register off112 = R28; |
mdoerr@9497 | 3554 | Register rIdx = R29; |
mdoerr@9497 | 3555 | Register rMax = R30; |
mdoerr@9497 | 3556 | Register constantsPos = R31; |
mdoerr@9497 | 3557 | |
mdoerr@9497 | 3558 | VectorRegister mask_32bit = VR24; |
mdoerr@9497 | 3559 | VectorRegister mask_64bit = VR25; |
mdoerr@9497 | 3560 | VectorRegister zeroes = VR26; |
mdoerr@9497 | 3561 | VectorRegister const1 = VR27; |
mdoerr@9497 | 3562 | VectorRegister const2 = VR28; |
mdoerr@9497 | 3563 | |
mdoerr@9497 | 3564 | // Save non-volatile vector registers (frameless). |
mdoerr@9497 | 3565 | Register offset = t1; int offsetInt = 0; |
mdoerr@9497 | 3566 | offsetInt -= 16; li(offset, -16); stvx(VR20, offset, R1_SP); |
mdoerr@9497 | 3567 | offsetInt -= 16; addi(offset, offset, -16); stvx(VR21, offset, R1_SP); |
mdoerr@9497 | 3568 | offsetInt -= 16; addi(offset, offset, -16); stvx(VR22, offset, R1_SP); |
mdoerr@9497 | 3569 | offsetInt -= 16; addi(offset, offset, -16); stvx(VR23, offset, R1_SP); |
mdoerr@9497 | 3570 | offsetInt -= 16; addi(offset, offset, -16); stvx(VR24, offset, R1_SP); |
mdoerr@9497 | 3571 | offsetInt -= 16; addi(offset, offset, -16); stvx(VR25, offset, R1_SP); |
mdoerr@9497 | 3572 | offsetInt -= 16; addi(offset, offset, -16); stvx(VR26, offset, R1_SP); |
mdoerr@9497 | 3573 | offsetInt -= 16; addi(offset, offset, -16); stvx(VR27, offset, R1_SP); |
mdoerr@9497 | 3574 | offsetInt -= 16; addi(offset, offset, -16); stvx(VR28, offset, R1_SP); |
mdoerr@9497 | 3575 | offsetInt -= 8; std(R22, offsetInt, R1_SP); |
mdoerr@9497 | 3576 | offsetInt -= 8; std(R23, offsetInt, R1_SP); |
mdoerr@9497 | 3577 | offsetInt -= 8; std(R24, offsetInt, R1_SP); |
mdoerr@9497 | 3578 | offsetInt -= 8; std(R25, offsetInt, R1_SP); |
mdoerr@9497 | 3579 | offsetInt -= 8; std(R26, offsetInt, R1_SP); |
mdoerr@9497 | 3580 | offsetInt -= 8; std(R27, offsetInt, R1_SP); |
mdoerr@9497 | 3581 | offsetInt -= 8; std(R28, offsetInt, R1_SP); |
mdoerr@9497 | 3582 | offsetInt -= 8; std(R29, offsetInt, R1_SP); |
mdoerr@9497 | 3583 | offsetInt -= 8; std(R30, offsetInt, R1_SP); |
mdoerr@9497 | 3584 | offsetInt -= 8; std(R31, offsetInt, R1_SP); |
mdoerr@9497 | 3585 | |
mdoerr@9497 | 3586 | // Set constants |
mdoerr@9497 | 3587 | li(off16, 16); |
mdoerr@9497 | 3588 | li(off32, 32); |
mdoerr@9497 | 3589 | li(off48, 48); |
mdoerr@9497 | 3590 | li(off64, 64); |
mdoerr@9497 | 3591 | li(off80, 80); |
mdoerr@9497 | 3592 | li(off96, 96); |
mdoerr@9497 | 3593 | li(off112, 112); |
mdoerr@9497 | 3594 | |
mdoerr@9497 | 3595 | clrldi(crc, crc, 32); |
mdoerr@9497 | 3596 | |
mdoerr@9497 | 3597 | vxor(zeroes, zeroes, zeroes); |
mdoerr@9497 | 3598 | vspltisw(VR0, -1); |
mdoerr@9497 | 3599 | |
mdoerr@9497 | 3600 | vsldoi(mask_32bit, zeroes, VR0, 4); |
mdoerr@9603 | 3601 | vsldoi(mask_64bit, zeroes, VR0, 8); |
mdoerr@9497 | 3602 | |
mdoerr@9497 | 3603 | // Get the initial value into v8 |
mdoerr@9497 | 3604 | vxor(VR8, VR8, VR8); |
mdoerr@9497 | 3605 | mtvrd(VR8, crc); |
mdoerr@9603 | 3606 | vsldoi(VR8, zeroes, VR8, 8); // shift into bottom 32 bits |
mdoerr@9497 | 3607 | |
mdoerr@9497 | 3608 | li (rLoaded, 0); |
mdoerr@9497 | 3609 | |
mdoerr@9497 | 3610 | rldicr(rIdx, len, 0, 56); |
mdoerr@9497 | 3611 | |
mdoerr@9497 | 3612 | { |
mdoerr@9497 | 3613 | BIND(L_1); |
mdoerr@9497 | 3614 | // Checksum in blocks of MAX_SIZE (32768) |
mdoerr@9497 | 3615 | lis(rMax, 0); |
mdoerr@9497 | 3616 | ori(rMax, rMax, 32768); |
mdoerr@9497 | 3617 | mr(rTmp2, rMax); |
mdoerr@9497 | 3618 | cmpd(CCR0, rIdx, rMax); |
mdoerr@9497 | 3619 | bgt(CCR0, L_2); |
mdoerr@9497 | 3620 | mr(rMax, rIdx); |
mdoerr@9497 | 3621 | |
mdoerr@9497 | 3622 | BIND(L_2); |
mdoerr@9497 | 3623 | subf(rIdx, rMax, rIdx); |
mdoerr@9497 | 3624 | |
mdoerr@9497 | 3625 | // our main loop does 128 bytes at a time |
mdoerr@9497 | 3626 | srdi(rMax, rMax, 7); |
mdoerr@9497 | 3627 | |
mdoerr@9497 | 3628 | /* |
mdoerr@9497 | 3629 | * Work out the offset into the constants table to start at. Each |
mdoerr@9497 | 3630 | * constant is 16 bytes, and it is used against 128 bytes of input |
mdoerr@9497 | 3631 | * data - 128 / 16 = 8 |
mdoerr@9497 | 3632 | */ |
mdoerr@9497 | 3633 | sldi(rTmp1, rMax, 4); |
mdoerr@9497 | 3634 | srdi(rTmp2, rTmp2, 3); |
mdoerr@9497 | 3635 | subf(rTmp1, rTmp1, rTmp2); |
mdoerr@9497 | 3636 | |
mdoerr@9497 | 3637 | // We reduce our final 128 bytes in a separate step |
mdoerr@9497 | 3638 | addi(rMax, rMax, -1); |
mdoerr@9497 | 3639 | mtctr(rMax); |
mdoerr@9497 | 3640 | |
mdoerr@9497 | 3641 | // Find the start of our constants |
mdoerr@9497 | 3642 | add(constantsPos, constants, rTmp1); |
mdoerr@9497 | 3643 | |
mdoerr@9497 | 3644 | // zero VR0-v7 which will contain our checksums |
mdoerr@9497 | 3645 | vxor(VR0, VR0, VR0); |
mdoerr@9497 | 3646 | vxor(VR1, VR1, VR1); |
mdoerr@9497 | 3647 | vxor(VR2, VR2, VR2); |
mdoerr@9497 | 3648 | vxor(VR3, VR3, VR3); |
mdoerr@9497 | 3649 | vxor(VR4, VR4, VR4); |
mdoerr@9497 | 3650 | vxor(VR5, VR5, VR5); |
mdoerr@9497 | 3651 | vxor(VR6, VR6, VR6); |
mdoerr@9497 | 3652 | vxor(VR7, VR7, VR7); |
mdoerr@9497 | 3653 | |
mdoerr@9497 | 3654 | lvx(const1, constantsPos); |
mdoerr@9497 | 3655 | |
mdoerr@9497 | 3656 | /* |
mdoerr@9497 | 3657 | * If we are looping back to consume more data we use the values |
mdoerr@9497 | 3658 | * already in VR16-v23. |
mdoerr@9497 | 3659 | */ |
mdoerr@9497 | 3660 | cmpdi(CCR0, rLoaded, 1); |
mdoerr@9497 | 3661 | beq(CCR0, L_3); |
mdoerr@9497 | 3662 | { |
mdoerr@9497 | 3663 | |
mdoerr@9497 | 3664 | // First warm up pass |
mdoerr@9497 | 3665 | lvx(VR16, buf); |
mdoerr@9497 | 3666 | lvx(VR17, off16, buf); |
mdoerr@9497 | 3667 | lvx(VR18, off32, buf); |
mdoerr@9497 | 3668 | lvx(VR19, off48, buf); |
mdoerr@9497 | 3669 | lvx(VR20, off64, buf); |
mdoerr@9497 | 3670 | lvx(VR21, off80, buf); |
mdoerr@9497 | 3671 | lvx(VR22, off96, buf); |
mdoerr@9497 | 3672 | lvx(VR23, off112, buf); |
mdoerr@9497 | 3673 | addi(buf, buf, 8*16); |
mdoerr@9497 | 3674 | |
mdoerr@9497 | 3675 | // xor in initial value |
mdoerr@9497 | 3676 | vxor(VR16, VR16, VR8); |
mdoerr@9497 | 3677 | } |
mdoerr@9497 | 3678 | |
mdoerr@9497 | 3679 | BIND(L_3); |
mdoerr@9497 | 3680 | bdz(L_first_warm_up_done); |
mdoerr@9497 | 3681 | |
mdoerr@9497 | 3682 | addi(constantsPos, constantsPos, 16); |
mdoerr@9497 | 3683 | lvx(const2, constantsPos); |
mdoerr@9497 | 3684 | |
mdoerr@9497 | 3685 | // Second warm up pass |
mdoerr@9497 | 3686 | vpmsumd(VR8, VR16, const1); |
mdoerr@9497 | 3687 | lvx(VR16, buf); |
mdoerr@9497 | 3688 | |
mdoerr@9497 | 3689 | vpmsumd(VR9, VR17, const1); |
mdoerr@9497 | 3690 | lvx(VR17, off16, buf); |
mdoerr@9497 | 3691 | |
mdoerr@9497 | 3692 | vpmsumd(VR10, VR18, const1); |
mdoerr@9497 | 3693 | lvx(VR18, off32, buf); |
mdoerr@9497 | 3694 | |
mdoerr@9497 | 3695 | vpmsumd(VR11, VR19, const1); |
mdoerr@9497 | 3696 | lvx(VR19, off48, buf); |
mdoerr@9497 | 3697 | |
mdoerr@9497 | 3698 | vpmsumd(VR12, VR20, const1); |
mdoerr@9497 | 3699 | lvx(VR20, off64, buf); |
mdoerr@9497 | 3700 | |
mdoerr@9497 | 3701 | vpmsumd(VR13, VR21, const1); |
mdoerr@9497 | 3702 | lvx(VR21, off80, buf); |
mdoerr@9497 | 3703 | |
mdoerr@9497 | 3704 | vpmsumd(VR14, VR22, const1); |
mdoerr@9497 | 3705 | lvx(VR22, off96, buf); |
mdoerr@9497 | 3706 | |
mdoerr@9497 | 3707 | vpmsumd(VR15, VR23, const1); |
mdoerr@9497 | 3708 | lvx(VR23, off112, buf); |
mdoerr@9497 | 3709 | |
mdoerr@9497 | 3710 | addi(buf, buf, 8 * 16); |
mdoerr@9497 | 3711 | |
mdoerr@9497 | 3712 | bdz(L_first_cool_down); |
mdoerr@9497 | 3713 | |
mdoerr@9497 | 3714 | /* |
mdoerr@9497 | 3715 | * main loop. We modulo schedule it such that it takes three iterations |
mdoerr@9497 | 3716 | * to complete - first iteration load, second iteration vpmsum, third |
mdoerr@9497 | 3717 | * iteration xor. |
mdoerr@9497 | 3718 | */ |
mdoerr@9497 | 3719 | { |
mdoerr@9497 | 3720 | BIND(L_4); |
mdoerr@9497 | 3721 | lvx(const1, constantsPos); addi(constantsPos, constantsPos, 16); |
mdoerr@9497 | 3722 | |
mdoerr@9497 | 3723 | vxor(VR0, VR0, VR8); |
mdoerr@9497 | 3724 | vpmsumd(VR8, VR16, const2); |
mdoerr@9497 | 3725 | lvx(VR16, buf); |
mdoerr@9497 | 3726 | |
mdoerr@9497 | 3727 | vxor(VR1, VR1, VR9); |
mdoerr@9497 | 3728 | vpmsumd(VR9, VR17, const2); |
mdoerr@9497 | 3729 | lvx(VR17, off16, buf); |
mdoerr@9497 | 3730 | |
mdoerr@9497 | 3731 | vxor(VR2, VR2, VR10); |
mdoerr@9497 | 3732 | vpmsumd(VR10, VR18, const2); |
mdoerr@9497 | 3733 | lvx(VR18, off32, buf); |
mdoerr@9497 | 3734 | |
mdoerr@9497 | 3735 | vxor(VR3, VR3, VR11); |
mdoerr@9497 | 3736 | vpmsumd(VR11, VR19, const2); |
mdoerr@9497 | 3737 | lvx(VR19, off48, buf); |
mdoerr@9497 | 3738 | lvx(const2, constantsPos); |
mdoerr@9497 | 3739 | |
mdoerr@9497 | 3740 | vxor(VR4, VR4, VR12); |
mdoerr@9497 | 3741 | vpmsumd(VR12, VR20, const1); |
mdoerr@9497 | 3742 | lvx(VR20, off64, buf); |
mdoerr@9497 | 3743 | |
mdoerr@9497 | 3744 | vxor(VR5, VR5, VR13); |
mdoerr@9497 | 3745 | vpmsumd(VR13, VR21, const1); |
mdoerr@9497 | 3746 | lvx(VR21, off80, buf); |
mdoerr@9497 | 3747 | |
mdoerr@9497 | 3748 | vxor(VR6, VR6, VR14); |
mdoerr@9497 | 3749 | vpmsumd(VR14, VR22, const1); |
mdoerr@9497 | 3750 | lvx(VR22, off96, buf); |
mdoerr@9497 | 3751 | |
mdoerr@9497 | 3752 | vxor(VR7, VR7, VR15); |
mdoerr@9497 | 3753 | vpmsumd(VR15, VR23, const1); |
mdoerr@9497 | 3754 | lvx(VR23, off112, buf); |
mdoerr@9497 | 3755 | |
mdoerr@9497 | 3756 | addi(buf, buf, 8 * 16); |
mdoerr@9497 | 3757 | |
mdoerr@9497 | 3758 | bdnz(L_4); |
mdoerr@9497 | 3759 | } |
mdoerr@9497 | 3760 | |
mdoerr@9497 | 3761 | BIND(L_first_cool_down); |
mdoerr@9497 | 3762 | |
mdoerr@9497 | 3763 | // First cool down pass |
mdoerr@9497 | 3764 | lvx(const1, constantsPos); |
mdoerr@9497 | 3765 | addi(constantsPos, constantsPos, 16); |
mdoerr@9497 | 3766 | |
mdoerr@9497 | 3767 | vxor(VR0, VR0, VR8); |
mdoerr@9497 | 3768 | vpmsumd(VR8, VR16, const1); |
mdoerr@9497 | 3769 | |
mdoerr@9497 | 3770 | vxor(VR1, VR1, VR9); |
mdoerr@9497 | 3771 | vpmsumd(VR9, VR17, const1); |
mdoerr@9497 | 3772 | |
mdoerr@9497 | 3773 | vxor(VR2, VR2, VR10); |
mdoerr@9497 | 3774 | vpmsumd(VR10, VR18, const1); |
mdoerr@9497 | 3775 | |
mdoerr@9497 | 3776 | vxor(VR3, VR3, VR11); |
mdoerr@9497 | 3777 | vpmsumd(VR11, VR19, const1); |
mdoerr@9497 | 3778 | |
mdoerr@9497 | 3779 | vxor(VR4, VR4, VR12); |
mdoerr@9497 | 3780 | vpmsumd(VR12, VR20, const1); |
mdoerr@9497 | 3781 | |
mdoerr@9497 | 3782 | vxor(VR5, VR5, VR13); |
mdoerr@9497 | 3783 | vpmsumd(VR13, VR21, const1); |
mdoerr@9497 | 3784 | |
mdoerr@9497 | 3785 | vxor(VR6, VR6, VR14); |
mdoerr@9497 | 3786 | vpmsumd(VR14, VR22, const1); |
mdoerr@9497 | 3787 | |
mdoerr@9497 | 3788 | vxor(VR7, VR7, VR15); |
mdoerr@9497 | 3789 | vpmsumd(VR15, VR23, const1); |
mdoerr@9497 | 3790 | |
mdoerr@9497 | 3791 | BIND(L_second_cool_down); |
mdoerr@9497 | 3792 | // Second cool down pass |
mdoerr@9497 | 3793 | vxor(VR0, VR0, VR8); |
mdoerr@9497 | 3794 | vxor(VR1, VR1, VR9); |
mdoerr@9497 | 3795 | vxor(VR2, VR2, VR10); |
mdoerr@9497 | 3796 | vxor(VR3, VR3, VR11); |
mdoerr@9497 | 3797 | vxor(VR4, VR4, VR12); |
mdoerr@9497 | 3798 | vxor(VR5, VR5, VR13); |
mdoerr@9497 | 3799 | vxor(VR6, VR6, VR14); |
mdoerr@9497 | 3800 | vxor(VR7, VR7, VR15); |
mdoerr@9497 | 3801 | |
mdoerr@9497 | 3802 | /* |
mdoerr@9497 | 3803 | * vpmsumd produces a 96 bit result in the least significant bits |
mdoerr@9497 | 3804 | * of the register. Since we are bit reflected we have to shift it |
mdoerr@9497 | 3805 | * left 32 bits so it occupies the least significant bits in the |
mdoerr@9497 | 3806 | * bit reflected domain. |
mdoerr@9497 | 3807 | */ |
mdoerr@9497 | 3808 | vsldoi(VR0, VR0, zeroes, 4); |
mdoerr@9497 | 3809 | vsldoi(VR1, VR1, zeroes, 4); |
mdoerr@9497 | 3810 | vsldoi(VR2, VR2, zeroes, 4); |
mdoerr@9497 | 3811 | vsldoi(VR3, VR3, zeroes, 4); |
mdoerr@9497 | 3812 | vsldoi(VR4, VR4, zeroes, 4); |
mdoerr@9497 | 3813 | vsldoi(VR5, VR5, zeroes, 4); |
mdoerr@9497 | 3814 | vsldoi(VR6, VR6, zeroes, 4); |
mdoerr@9497 | 3815 | vsldoi(VR7, VR7, zeroes, 4); |
mdoerr@9497 | 3816 | |
mdoerr@9497 | 3817 | // xor with last 1024 bits |
mdoerr@9497 | 3818 | lvx(VR8, buf); |
mdoerr@9497 | 3819 | lvx(VR9, off16, buf); |
mdoerr@9497 | 3820 | lvx(VR10, off32, buf); |
mdoerr@9497 | 3821 | lvx(VR11, off48, buf); |
mdoerr@9497 | 3822 | lvx(VR12, off64, buf); |
mdoerr@9497 | 3823 | lvx(VR13, off80, buf); |
mdoerr@9497 | 3824 | lvx(VR14, off96, buf); |
mdoerr@9497 | 3825 | lvx(VR15, off112, buf); |
mdoerr@9497 | 3826 | addi(buf, buf, 8 * 16); |
mdoerr@9497 | 3827 | |
mdoerr@9497 | 3828 | vxor(VR16, VR0, VR8); |
mdoerr@9497 | 3829 | vxor(VR17, VR1, VR9); |
mdoerr@9497 | 3830 | vxor(VR18, VR2, VR10); |
mdoerr@9497 | 3831 | vxor(VR19, VR3, VR11); |
mdoerr@9497 | 3832 | vxor(VR20, VR4, VR12); |
mdoerr@9497 | 3833 | vxor(VR21, VR5, VR13); |
mdoerr@9497 | 3834 | vxor(VR22, VR6, VR14); |
mdoerr@9497 | 3835 | vxor(VR23, VR7, VR15); |
mdoerr@9497 | 3836 | |
mdoerr@9497 | 3837 | li(rLoaded, 1); |
mdoerr@9497 | 3838 | cmpdi(CCR0, rIdx, 0); |
mdoerr@9497 | 3839 | addi(rIdx, rIdx, 128); |
mdoerr@9497 | 3840 | bne(CCR0, L_1); |
mdoerr@9497 | 3841 | } |
mdoerr@9497 | 3842 | |
mdoerr@9497 | 3843 | // Work out how many bytes we have left |
mdoerr@9497 | 3844 | andi_(len, len, 127); |
mdoerr@9497 | 3845 | |
mdoerr@9497 | 3846 | // Calculate where in the constant table we need to start |
mdoerr@9497 | 3847 | subfic(rTmp1, len, 128); |
mdoerr@9497 | 3848 | add(constantsPos, constantsPos, rTmp1); |
mdoerr@9497 | 3849 | |
mdoerr@9497 | 3850 | // How many 16 byte chunks are in the tail |
mdoerr@9497 | 3851 | srdi(rIdx, len, 4); |
mdoerr@9497 | 3852 | mtctr(rIdx); |
mdoerr@9497 | 3853 | |
mdoerr@9497 | 3854 | /* |
mdoerr@9497 | 3855 | * Reduce the previously calculated 1024 bits to 64 bits, shifting |
mdoerr@9497 | 3856 | * 32 bits to include the trailing 32 bits of zeros |
mdoerr@9497 | 3857 | */ |
mdoerr@9497 | 3858 | lvx(VR0, constantsPos); |
mdoerr@9497 | 3859 | lvx(VR1, off16, constantsPos); |
mdoerr@9497 | 3860 | lvx(VR2, off32, constantsPos); |
mdoerr@9497 | 3861 | lvx(VR3, off48, constantsPos); |
mdoerr@9497 | 3862 | lvx(VR4, off64, constantsPos); |
mdoerr@9497 | 3863 | lvx(VR5, off80, constantsPos); |
mdoerr@9497 | 3864 | lvx(VR6, off96, constantsPos); |
mdoerr@9497 | 3865 | lvx(VR7, off112, constantsPos); |
mdoerr@9497 | 3866 | addi(constantsPos, constantsPos, 8 * 16); |
mdoerr@9497 | 3867 | |
mdoerr@9497 | 3868 | vpmsumw(VR0, VR16, VR0); |
mdoerr@9497 | 3869 | vpmsumw(VR1, VR17, VR1); |
mdoerr@9497 | 3870 | vpmsumw(VR2, VR18, VR2); |
mdoerr@9497 | 3871 | vpmsumw(VR3, VR19, VR3); |
mdoerr@9497 | 3872 | vpmsumw(VR4, VR20, VR4); |
mdoerr@9497 | 3873 | vpmsumw(VR5, VR21, VR5); |
mdoerr@9497 | 3874 | vpmsumw(VR6, VR22, VR6); |
mdoerr@9497 | 3875 | vpmsumw(VR7, VR23, VR7); |
mdoerr@9497 | 3876 | |
mdoerr@9497 | 3877 | // Now reduce the tail (0 - 112 bytes) |
mdoerr@9497 | 3878 | cmpdi(CCR0, rIdx, 0); |
mdoerr@9497 | 3879 | beq(CCR0, L_XOR); |
mdoerr@9497 | 3880 | |
mdoerr@9497 | 3881 | lvx(VR16, buf); addi(buf, buf, 16); |
mdoerr@9497 | 3882 | lvx(VR17, constantsPos); |
mdoerr@9497 | 3883 | vpmsumw(VR16, VR16, VR17); |
mdoerr@9497 | 3884 | vxor(VR0, VR0, VR16); |
mdoerr@9497 | 3885 | beq(CCR0, L_XOR); |
mdoerr@9497 | 3886 | |
mdoerr@9497 | 3887 | lvx(VR16, buf); addi(buf, buf, 16); |
mdoerr@9497 | 3888 | lvx(VR17, off16, constantsPos); |
mdoerr@9497 | 3889 | vpmsumw(VR16, VR16, VR17); |
mdoerr@9497 | 3890 | vxor(VR0, VR0, VR16); |
mdoerr@9497 | 3891 | beq(CCR0, L_XOR); |
mdoerr@9497 | 3892 | |
mdoerr@9497 | 3893 | lvx(VR16, buf); addi(buf, buf, 16); |
mdoerr@9497 | 3894 | lvx(VR17, off32, constantsPos); |
mdoerr@9497 | 3895 | vpmsumw(VR16, VR16, VR17); |
mdoerr@9497 | 3896 | vxor(VR0, VR0, VR16); |
mdoerr@9497 | 3897 | beq(CCR0, L_XOR); |
mdoerr@9497 | 3898 | |
mdoerr@9497 | 3899 | lvx(VR16, buf); addi(buf, buf, 16); |
mdoerr@9497 | 3900 | lvx(VR17, off48,constantsPos); |
mdoerr@9497 | 3901 | vpmsumw(VR16, VR16, VR17); |
mdoerr@9497 | 3902 | vxor(VR0, VR0, VR16); |
mdoerr@9497 | 3903 | beq(CCR0, L_XOR); |
mdoerr@9497 | 3904 | |
mdoerr@9497 | 3905 | lvx(VR16, buf); addi(buf, buf, 16); |
mdoerr@9497 | 3906 | lvx(VR17, off64, constantsPos); |
mdoerr@9497 | 3907 | vpmsumw(VR16, VR16, VR17); |
mdoerr@9497 | 3908 | vxor(VR0, VR0, VR16); |
mdoerr@9497 | 3909 | beq(CCR0, L_XOR); |
mdoerr@9497 | 3910 | |
mdoerr@9497 | 3911 | lvx(VR16, buf); addi(buf, buf, 16); |
mdoerr@9497 | 3912 | lvx(VR17, off80, constantsPos); |
mdoerr@9497 | 3913 | vpmsumw(VR16, VR16, VR17); |
mdoerr@9497 | 3914 | vxor(VR0, VR0, VR16); |
mdoerr@9497 | 3915 | beq(CCR0, L_XOR); |
mdoerr@9497 | 3916 | |
mdoerr@9497 | 3917 | lvx(VR16, buf); addi(buf, buf, 16); |
mdoerr@9497 | 3918 | lvx(VR17, off96, constantsPos); |
mdoerr@9497 | 3919 | vpmsumw(VR16, VR16, VR17); |
mdoerr@9497 | 3920 | vxor(VR0, VR0, VR16); |
mdoerr@9497 | 3921 | |
mdoerr@9497 | 3922 | // Now xor all the parallel chunks together |
mdoerr@9497 | 3923 | BIND(L_XOR); |
mdoerr@9497 | 3924 | vxor(VR0, VR0, VR1); |
mdoerr@9497 | 3925 | vxor(VR2, VR2, VR3); |
mdoerr@9497 | 3926 | vxor(VR4, VR4, VR5); |
mdoerr@9497 | 3927 | vxor(VR6, VR6, VR7); |
mdoerr@9497 | 3928 | |
mdoerr@9497 | 3929 | vxor(VR0, VR0, VR2); |
mdoerr@9497 | 3930 | vxor(VR4, VR4, VR6); |
mdoerr@9497 | 3931 | |
mdoerr@9497 | 3932 | vxor(VR0, VR0, VR4); |
mdoerr@9497 | 3933 | |
mdoerr@9497 | 3934 | b(L_barrett_reduction); |
mdoerr@9497 | 3935 | |
mdoerr@9497 | 3936 | BIND(L_first_warm_up_done); |
mdoerr@9497 | 3937 | lvx(const1, constantsPos); |
mdoerr@9497 | 3938 | addi(constantsPos, constantsPos, 16); |
mdoerr@9497 | 3939 | vpmsumd(VR8, VR16, const1); |
mdoerr@9497 | 3940 | vpmsumd(VR9, VR17, const1); |
mdoerr@9497 | 3941 | vpmsumd(VR10, VR18, const1); |
mdoerr@9497 | 3942 | vpmsumd(VR11, VR19, const1); |
mdoerr@9497 | 3943 | vpmsumd(VR12, VR20, const1); |
mdoerr@9497 | 3944 | vpmsumd(VR13, VR21, const1); |
mdoerr@9497 | 3945 | vpmsumd(VR14, VR22, const1); |
mdoerr@9497 | 3946 | vpmsumd(VR15, VR23, const1); |
mdoerr@9497 | 3947 | b(L_second_cool_down); |
mdoerr@9497 | 3948 | |
mdoerr@9497 | 3949 | BIND(L_barrett_reduction); |
mdoerr@9497 | 3950 | |
mdoerr@9497 | 3951 | lvx(const1, barretConstants); |
mdoerr@9497 | 3952 | addi(barretConstants, barretConstants, 16); |
mdoerr@9497 | 3953 | lvx(const2, barretConstants); |
mdoerr@9497 | 3954 | |
mdoerr@9603 | 3955 | vsldoi(VR1, VR0, VR0, 8); |
mdoerr@9497 | 3956 | vxor(VR0, VR0, VR1); // xor two 64 bit results together |
mdoerr@9497 | 3957 | |
mdoerr@9497 | 3958 | // shift left one bit |
mdoerr@9497 | 3959 | vspltisb(VR1, 1); |
mdoerr@9497 | 3960 | vsl(VR0, VR0, VR1); |
mdoerr@9497 | 3961 | |
mdoerr@9497 | 3962 | vand(VR0, VR0, mask_64bit); |
mdoerr@9497 | 3963 | |
mdoerr@9497 | 3964 | /* |
mdoerr@9497 | 3965 | * The reflected version of Barrett reduction. Instead of bit |
mdoerr@9497 | 3966 | * reflecting our data (which is expensive to do), we bit reflect our |
mdoerr@9497 | 3967 | * constants and our algorithm, which means the intermediate data in |
mdoerr@9497 | 3968 | * our vector registers goes from 0-63 instead of 63-0. We can reflect |
mdoerr@9497 | 3969 | * the algorithm because we don't carry in mod 2 arithmetic. |
mdoerr@9497 | 3970 | */ |
mdoerr@9497 | 3971 | vand(VR1, VR0, mask_32bit); // bottom 32 bits of a |
mdoerr@9497 | 3972 | vpmsumd(VR1, VR1, const1); // ma |
mdoerr@9497 | 3973 | vand(VR1, VR1, mask_32bit); // bottom 32bits of ma |
mdoerr@9497 | 3974 | vpmsumd(VR1, VR1, const2); // qn */ |
mdoerr@9497 | 3975 | vxor(VR0, VR0, VR1); // a - qn, subtraction is xor in GF(2) |
mdoerr@9497 | 3976 | |
mdoerr@9497 | 3977 | /* |
mdoerr@9497 | 3978 | * Since we are bit reflected, the result (ie the low 32 bits) is in |
mdoerr@9497 | 3979 | * the high 32 bits. We just need to shift it left 4 bytes |
mdoerr@9497 | 3980 | * V0 [ 0 1 X 3 ] |
mdoerr@9497 | 3981 | * V0 [ 0 X 2 3 ] |
mdoerr@9497 | 3982 | */ |
mdoerr@9497 | 3983 | vsldoi(VR0, VR0, zeroes, 4); // shift result into top 64 bits of |
mdoerr@9497 | 3984 | |
mdoerr@9497 | 3985 | // Get it into r3 |
mdoerr@9497 | 3986 | mfvrd(crc, VR0); |
mdoerr@9497 | 3987 | |
mdoerr@9497 | 3988 | BIND(L_end); |
mdoerr@9497 | 3989 | |
mdoerr@9497 | 3990 | offsetInt = 0; |
mdoerr@9497 | 3991 | // Restore non-volatile Vector registers (frameless). |
mdoerr@9497 | 3992 | offsetInt -= 16; li(offset, -16); lvx(VR20, offset, R1_SP); |
mdoerr@9497 | 3993 | offsetInt -= 16; addi(offset, offset, -16); lvx(VR21, offset, R1_SP); |
mdoerr@9497 | 3994 | offsetInt -= 16; addi(offset, offset, -16); lvx(VR22, offset, R1_SP); |
mdoerr@9497 | 3995 | offsetInt -= 16; addi(offset, offset, -16); lvx(VR23, offset, R1_SP); |
mdoerr@9497 | 3996 | offsetInt -= 16; addi(offset, offset, -16); lvx(VR24, offset, R1_SP); |
mdoerr@9497 | 3997 | offsetInt -= 16; addi(offset, offset, -16); lvx(VR25, offset, R1_SP); |
mdoerr@9497 | 3998 | offsetInt -= 16; addi(offset, offset, -16); lvx(VR26, offset, R1_SP); |
mdoerr@9497 | 3999 | offsetInt -= 16; addi(offset, offset, -16); lvx(VR27, offset, R1_SP); |
mdoerr@9497 | 4000 | offsetInt -= 16; addi(offset, offset, -16); lvx(VR28, offset, R1_SP); |
mdoerr@9497 | 4001 | offsetInt -= 8; ld(R22, offsetInt, R1_SP); |
mdoerr@9497 | 4002 | offsetInt -= 8; ld(R23, offsetInt, R1_SP); |
mdoerr@9497 | 4003 | offsetInt -= 8; ld(R24, offsetInt, R1_SP); |
mdoerr@9497 | 4004 | offsetInt -= 8; ld(R25, offsetInt, R1_SP); |
mdoerr@9497 | 4005 | offsetInt -= 8; ld(R26, offsetInt, R1_SP); |
mdoerr@9497 | 4006 | offsetInt -= 8; ld(R27, offsetInt, R1_SP); |
mdoerr@9497 | 4007 | offsetInt -= 8; ld(R28, offsetInt, R1_SP); |
mdoerr@9497 | 4008 | offsetInt -= 8; ld(R29, offsetInt, R1_SP); |
mdoerr@9497 | 4009 | offsetInt -= 8; ld(R30, offsetInt, R1_SP); |
mdoerr@9497 | 4010 | offsetInt -= 8; ld(R31, offsetInt, R1_SP); |
mdoerr@9497 | 4011 | } |
mdoerr@9497 | 4012 | |
gromero@9496 | 4013 | void MacroAssembler::kernel_crc32_singleByte(Register crc, Register buf, Register len, Register table, Register tmp) { |
gromero@9496 | 4014 | assert_different_registers(crc, buf, /* len, not used!! */ table, tmp); |
gromero@9496 | 4015 | |
gromero@9496 | 4016 | BLOCK_COMMENT("kernel_crc32_singleByte:"); |
gromero@9496 | 4017 | nand(crc, crc, crc); // ~c |
gromero@9496 | 4018 | |
gromero@9496 | 4019 | lbz(tmp, 0, buf); // Byte from buffer, zero-extended. |
gromero@9496 | 4020 | update_byte_crc32(crc, tmp, table); |
gromero@9496 | 4021 | |
gromero@9496 | 4022 | nand(crc, crc, crc); // ~c |
gromero@9496 | 4023 | } |
gromero@9496 | 4024 | |
goetz@6458 | 4025 | |
goetz@6458 | 4026 | void MacroAssembler::asm_assert(bool check_equal, const char *msg, int id) { |
goetz@6458 | 4027 | #ifdef ASSERT |
goetz@6458 | 4028 | Label ok; |
goetz@6458 | 4029 | if (check_equal) { |
goetz@6458 | 4030 | beq(CCR0, ok); |
goetz@6458 | 4031 | } else { |
goetz@6458 | 4032 | bne(CCR0, ok); |
goetz@6458 | 4033 | } |
goetz@6458 | 4034 | stop(msg, id); |
goetz@6458 | 4035 | bind(ok); |
goetz@6458 | 4036 | #endif |
goetz@6458 | 4037 | } |
goetz@6458 | 4038 | |
goetz@6458 | 4039 | void MacroAssembler::asm_assert_mems_zero(bool check_equal, int size, int mem_offset, |
goetz@6458 | 4040 | Register mem_base, const char* msg, int id) { |
goetz@6458 | 4041 | #ifdef ASSERT |
goetz@6458 | 4042 | switch (size) { |
goetz@6458 | 4043 | case 4: |
goetz@6458 | 4044 | lwz(R0, mem_offset, mem_base); |
goetz@6458 | 4045 | cmpwi(CCR0, R0, 0); |
goetz@6458 | 4046 | break; |
goetz@6458 | 4047 | case 8: |
goetz@6458 | 4048 | ld(R0, mem_offset, mem_base); |
goetz@6458 | 4049 | cmpdi(CCR0, R0, 0); |
goetz@6458 | 4050 | break; |
goetz@6458 | 4051 | default: |
goetz@6458 | 4052 | ShouldNotReachHere(); |
goetz@6458 | 4053 | } |
goetz@6458 | 4054 | asm_assert(check_equal, msg, id); |
goetz@6458 | 4055 | #endif // ASSERT |
goetz@6458 | 4056 | } |
goetz@6458 | 4057 | |
goetz@6458 | 4058 | void MacroAssembler::verify_thread() { |
goetz@6458 | 4059 | if (VerifyThread) { |
goetz@6458 | 4060 | unimplemented("'VerifyThread' currently not implemented on PPC"); |
goetz@6458 | 4061 | } |
goetz@6458 | 4062 | } |
goetz@6458 | 4063 | |
goetz@6458 | 4064 | // READ: oop. KILL: R0. Volatile floats perhaps. |
goetz@6458 | 4065 | void MacroAssembler::verify_oop(Register oop, const char* msg) { |
goetz@6458 | 4066 | if (!VerifyOops) { |
goetz@6458 | 4067 | return; |
goetz@6458 | 4068 | } |
goetz@7424 | 4069 | |
goetz@6495 | 4070 | address/* FunctionDescriptor** */fd = StubRoutines::verify_oop_subroutine_entry_address(); |
goetz@7424 | 4071 | const Register tmp = R11; // Will be preserved. |
goetz@7424 | 4072 | const int nbytes_save = 11*8; // Volatile gprs except R0. |
goetz@7424 | 4073 | save_volatile_gprs(R1_SP, -nbytes_save); // except R0 |
goetz@7424 | 4074 | |
goetz@7424 | 4075 | if (oop == tmp) mr(R4_ARG2, oop); |
goetz@7424 | 4076 | save_LR_CR(tmp); // save in old frame |
goetz@6511 | 4077 | push_frame_reg_args(nbytes_save, tmp); |
goetz@6511 | 4078 | // load FunctionDescriptor** / entry_address * |
goetz@7424 | 4079 | load_const_optimized(tmp, fd, R0); |
goetz@6511 | 4080 | // load FunctionDescriptor* / entry_address |
goetz@6458 | 4081 | ld(tmp, 0, tmp); |
goetz@7424 | 4082 | if (oop != tmp) mr_if_needed(R4_ARG2, oop); |
goetz@7424 | 4083 | load_const_optimized(R3_ARG1, (address)msg, R0); |
goetz@7424 | 4084 | // Call destination for its side effect. |
goetz@6458 | 4085 | call_c(tmp); |
goetz@7424 | 4086 | |
goetz@6458 | 4087 | pop_frame(); |
goetz@6458 | 4088 | restore_LR_CR(tmp); |
goetz@7424 | 4089 | restore_volatile_gprs(R1_SP, -nbytes_save); // except R0 |
goetz@6458 | 4090 | } |
goetz@6458 | 4091 | |
goetz@6458 | 4092 | const char* stop_types[] = { |
goetz@6458 | 4093 | "stop", |
goetz@6458 | 4094 | "untested", |
goetz@6458 | 4095 | "unimplemented", |
goetz@6458 | 4096 | "shouldnotreachhere" |
goetz@6458 | 4097 | }; |
goetz@6458 | 4098 | |
goetz@6458 | 4099 | static void stop_on_request(int tp, const char* msg) { |
coleenp@7358 | 4100 | tty->print("PPC assembly code requires stop: (%s) %s\n", stop_types[tp%/*stop_end*/4], msg); |
goetz@6458 | 4101 | guarantee(false, err_msg("PPC assembly code requires stop: %s", msg)); |
goetz@6458 | 4102 | } |
goetz@6458 | 4103 | |
goetz@6458 | 4104 | // Call a C-function that prints output. |
goetz@6458 | 4105 | void MacroAssembler::stop(int type, const char* msg, int id) { |
goetz@6458 | 4106 | #ifndef PRODUCT |
goetz@6458 | 4107 | block_comment(err_msg("stop: %s %s {", stop_types[type%stop_end], msg)); |
goetz@6458 | 4108 | #else |
goetz@6458 | 4109 | block_comment("stop {"); |
goetz@6458 | 4110 | #endif |
goetz@6458 | 4111 | |
goetz@6458 | 4112 | // setup arguments |
goetz@6458 | 4113 | load_const_optimized(R3_ARG1, type); |
goetz@6458 | 4114 | load_const_optimized(R4_ARG2, (void *)msg, /*tmp=*/R0); |
goetz@6458 | 4115 | call_VM_leaf(CAST_FROM_FN_PTR(address, stop_on_request), R3_ARG1, R4_ARG2); |
goetz@6458 | 4116 | illtrap(); |
goetz@6458 | 4117 | emit_int32(id); |
goetz@6458 | 4118 | block_comment("} stop;"); |
goetz@6458 | 4119 | } |
goetz@6458 | 4120 | |
goetz@6458 | 4121 | #ifndef PRODUCT |
goetz@6458 | 4122 | // Write pattern 0x0101010101010101 in memory region [low-before, high+after]. |
goetz@6458 | 4123 | // Val, addr are temp registers. |
goetz@6458 | 4124 | // If low == addr, addr is killed. |
goetz@6458 | 4125 | // High is preserved. |
goetz@6458 | 4126 | void MacroAssembler::zap_from_to(Register low, int before, Register high, int after, Register val, Register addr) { |
goetz@6458 | 4127 | if (!ZapMemory) return; |
goetz@6458 | 4128 | |
goetz@6458 | 4129 | assert_different_registers(low, val); |
goetz@6458 | 4130 | |
goetz@6458 | 4131 | BLOCK_COMMENT("zap memory region {"); |
goetz@6458 | 4132 | load_const_optimized(val, 0x0101010101010101); |
goetz@6458 | 4133 | int size = before + after; |
goetz@6458 | 4134 | if (low == high && size < 5 && size > 0) { |
goetz@6458 | 4135 | int offset = -before*BytesPerWord; |
goetz@6458 | 4136 | for (int i = 0; i < size; ++i) { |
goetz@6458 | 4137 | std(val, offset, low); |
goetz@6458 | 4138 | offset += (1*BytesPerWord); |
goetz@6458 | 4139 | } |
goetz@6458 | 4140 | } else { |
goetz@6458 | 4141 | addi(addr, low, -before*BytesPerWord); |
goetz@6458 | 4142 | assert_different_registers(high, val); |
goetz@6458 | 4143 | if (after) addi(high, high, after * BytesPerWord); |
goetz@6458 | 4144 | Label loop; |
goetz@6458 | 4145 | bind(loop); |
goetz@6458 | 4146 | std(val, 0, addr); |
goetz@6458 | 4147 | addi(addr, addr, 8); |
goetz@6458 | 4148 | cmpd(CCR6, addr, high); |
goetz@6458 | 4149 | ble(CCR6, loop); |
goetz@6458 | 4150 | if (after) addi(high, high, -after * BytesPerWord); // Correct back to old value. |
goetz@6458 | 4151 | } |
goetz@6458 | 4152 | BLOCK_COMMENT("} zap memory region"); |
goetz@6458 | 4153 | } |
goetz@6458 | 4154 | |
goetz@6458 | 4155 | #endif // !PRODUCT |
goetz@6512 | 4156 | |
goetz@6512 | 4157 | SkipIfEqualZero::SkipIfEqualZero(MacroAssembler* masm, Register temp, const bool* flag_addr) : _masm(masm), _label() { |
goetz@6512 | 4158 | int simm16_offset = masm->load_const_optimized(temp, (address)flag_addr, R0, true); |
goetz@6512 | 4159 | assert(sizeof(bool) == 1, "PowerPC ABI"); |
goetz@6512 | 4160 | masm->lbz(temp, simm16_offset, temp); |
goetz@6512 | 4161 | masm->cmpwi(CCR0, temp, 0); |
goetz@6512 | 4162 | masm->beq(CCR0, _label); |
goetz@6512 | 4163 | } |
goetz@6512 | 4164 | |
goetz@6512 | 4165 | SkipIfEqualZero::~SkipIfEqualZero() { |
goetz@6512 | 4166 | _masm->bind(_label); |
goetz@6512 | 4167 | } |