1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000 1.2 +++ b/src/cpu/x86/vm/c1_MacroAssembler_x86.cpp Sat Dec 01 00:00:00 2007 +0000 1.3 @@ -0,0 +1,385 @@ 1.4 +/* 1.5 + * Copyright 1999-2007 Sun Microsystems, Inc. All Rights Reserved. 1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 1.7 + * 1.8 + * This code is free software; you can redistribute it and/or modify it 1.9 + * under the terms of the GNU General Public License version 2 only, as 1.10 + * published by the Free Software Foundation. 1.11 + * 1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT 1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 1.15 + * version 2 for more details (a copy is included in the LICENSE file that 1.16 + * accompanied this code). 1.17 + * 1.18 + * You should have received a copy of the GNU General Public License version 1.19 + * 2 along with this work; if not, write to the Free Software Foundation, 1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 1.21 + * 1.22 + * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 1.23 + * CA 95054 USA or visit www.sun.com if you need additional information or 1.24 + * have any questions. 1.25 + * 1.26 + */ 1.27 + 1.28 +#include "incls/_precompiled.incl" 1.29 +#include "incls/_c1_MacroAssembler_x86.cpp.incl" 1.30 + 1.31 +int C1_MacroAssembler::lock_object(Register hdr, Register obj, Register disp_hdr, Register scratch, Label& slow_case) { 1.32 + const int aligned_mask = 3; 1.33 + const int hdr_offset = oopDesc::mark_offset_in_bytes(); 1.34 + assert(hdr == rax, "hdr must be rax, for the cmpxchg instruction"); 1.35 + assert(hdr != obj && hdr != disp_hdr && obj != disp_hdr, "registers must be different"); 1.36 + assert(BytesPerWord == 4, "adjust aligned_mask and code"); 1.37 + Label done; 1.38 + int null_check_offset = -1; 1.39 + 1.40 + verify_oop(obj); 1.41 + 1.42 + // save object being locked into the BasicObjectLock 1.43 + movl(Address(disp_hdr, BasicObjectLock::obj_offset_in_bytes()), obj); 1.44 + 1.45 + if (UseBiasedLocking) { 1.46 + assert(scratch != noreg, "should have scratch register at this point"); 1.47 + null_check_offset = biased_locking_enter(disp_hdr, obj, hdr, scratch, false, done, &slow_case); 1.48 + } else { 1.49 + null_check_offset = offset(); 1.50 + } 1.51 + 1.52 + // Load object header 1.53 + movl(hdr, Address(obj, hdr_offset)); 1.54 + // and mark it as unlocked 1.55 + orl(hdr, markOopDesc::unlocked_value); 1.56 + // save unlocked object header into the displaced header location on the stack 1.57 + movl(Address(disp_hdr, 0), hdr); 1.58 + // test if object header is still the same (i.e. unlocked), and if so, store the 1.59 + // displaced header address in the object header - if it is not the same, get the 1.60 + // object header instead 1.61 + if (os::is_MP()) MacroAssembler::lock(); // must be immediately before cmpxchg! 1.62 + cmpxchg(disp_hdr, Address(obj, hdr_offset)); 1.63 + // if the object header was the same, we're done 1.64 + if (PrintBiasedLockingStatistics) { 1.65 + cond_inc32(Assembler::equal, 1.66 + ExternalAddress((address)BiasedLocking::fast_path_entry_count_addr())); 1.67 + } 1.68 + jcc(Assembler::equal, done); 1.69 + // if the object header was not the same, it is now in the hdr register 1.70 + // => test if it is a stack pointer into the same stack (recursive locking), i.e.: 1.71 + // 1.72 + // 1) (hdr & aligned_mask) == 0 1.73 + // 2) rsp <= hdr 1.74 + // 3) hdr <= rsp + page_size 1.75 + // 1.76 + // these 3 tests can be done by evaluating the following expression: 1.77 + // 1.78 + // (hdr - rsp) & (aligned_mask - page_size) 1.79 + // 1.80 + // assuming both the stack pointer and page_size have their least 1.81 + // significant 2 bits cleared and page_size is a power of 2 1.82 + subl(hdr, rsp); 1.83 + andl(hdr, aligned_mask - os::vm_page_size()); 1.84 + // for recursive locking, the result is zero => save it in the displaced header 1.85 + // location (NULL in the displaced hdr location indicates recursive locking) 1.86 + movl(Address(disp_hdr, 0), hdr); 1.87 + // otherwise we don't care about the result and handle locking via runtime call 1.88 + jcc(Assembler::notZero, slow_case); 1.89 + // done 1.90 + bind(done); 1.91 + return null_check_offset; 1.92 +} 1.93 + 1.94 + 1.95 +void C1_MacroAssembler::unlock_object(Register hdr, Register obj, Register disp_hdr, Label& slow_case) { 1.96 + const int aligned_mask = 3; 1.97 + const int hdr_offset = oopDesc::mark_offset_in_bytes(); 1.98 + assert(disp_hdr == rax, "disp_hdr must be rax, for the cmpxchg instruction"); 1.99 + assert(hdr != obj && hdr != disp_hdr && obj != disp_hdr, "registers must be different"); 1.100 + assert(BytesPerWord == 4, "adjust aligned_mask and code"); 1.101 + Label done; 1.102 + 1.103 + if (UseBiasedLocking) { 1.104 + // load object 1.105 + movl(obj, Address(disp_hdr, BasicObjectLock::obj_offset_in_bytes())); 1.106 + biased_locking_exit(obj, hdr, done); 1.107 + } 1.108 + 1.109 + // load displaced header 1.110 + movl(hdr, Address(disp_hdr, 0)); 1.111 + // if the loaded hdr is NULL we had recursive locking 1.112 + testl(hdr, hdr); 1.113 + // if we had recursive locking, we are done 1.114 + jcc(Assembler::zero, done); 1.115 + if (!UseBiasedLocking) { 1.116 + // load object 1.117 + movl(obj, Address(disp_hdr, BasicObjectLock::obj_offset_in_bytes())); 1.118 + } 1.119 + verify_oop(obj); 1.120 + // test if object header is pointing to the displaced header, and if so, restore 1.121 + // the displaced header in the object - if the object header is not pointing to 1.122 + // the displaced header, get the object header instead 1.123 + if (os::is_MP()) MacroAssembler::lock(); // must be immediately before cmpxchg! 1.124 + cmpxchg(hdr, Address(obj, hdr_offset)); 1.125 + // if the object header was not pointing to the displaced header, 1.126 + // we do unlocking via runtime call 1.127 + jcc(Assembler::notEqual, slow_case); 1.128 + // done 1.129 + bind(done); 1.130 +} 1.131 + 1.132 + 1.133 +// Defines obj, preserves var_size_in_bytes 1.134 +void C1_MacroAssembler::try_allocate(Register obj, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2, Label& slow_case) { 1.135 + if (UseTLAB) { 1.136 + tlab_allocate(obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case); 1.137 + } else { 1.138 + eden_allocate(obj, var_size_in_bytes, con_size_in_bytes, t1, slow_case); 1.139 + } 1.140 +} 1.141 + 1.142 + 1.143 +void C1_MacroAssembler::initialize_header(Register obj, Register klass, Register len, Register t1, Register t2) { 1.144 + assert_different_registers(obj, klass, len); 1.145 + if (UseBiasedLocking && !len->is_valid()) { 1.146 + assert_different_registers(obj, klass, len, t1, t2); 1.147 + movl(t1, Address(klass, Klass::prototype_header_offset_in_bytes() + klassOopDesc::klass_part_offset_in_bytes())); 1.148 + movl(Address(obj, oopDesc::mark_offset_in_bytes()), t1); 1.149 + } else { 1.150 + movl(Address(obj, oopDesc::mark_offset_in_bytes ()), (int)markOopDesc::prototype()); 1.151 + } 1.152 + 1.153 + movl(Address(obj, oopDesc::klass_offset_in_bytes()), klass); 1.154 + if (len->is_valid()) { 1.155 + movl(Address(obj, arrayOopDesc::length_offset_in_bytes()), len); 1.156 + } 1.157 +} 1.158 + 1.159 + 1.160 +// preserves obj, destroys len_in_bytes 1.161 +void C1_MacroAssembler::initialize_body(Register obj, Register len_in_bytes, int hdr_size_in_bytes, Register t1) { 1.162 + Label done; 1.163 + assert(obj != len_in_bytes && obj != t1 && t1 != len_in_bytes, "registers must be different"); 1.164 + assert((hdr_size_in_bytes & (BytesPerWord - 1)) == 0, "header size is not a multiple of BytesPerWord"); 1.165 + Register index = len_in_bytes; 1.166 + subl(index, hdr_size_in_bytes); 1.167 + jcc(Assembler::zero, done); 1.168 + // initialize topmost word, divide index by 2, check if odd and test if zero 1.169 + // note: for the remaining code to work, index must be a multiple of BytesPerWord 1.170 +#ifdef ASSERT 1.171 + { Label L; 1.172 + testl(index, BytesPerWord - 1); 1.173 + jcc(Assembler::zero, L); 1.174 + stop("index is not a multiple of BytesPerWord"); 1.175 + bind(L); 1.176 + } 1.177 +#endif 1.178 + xorl(t1, t1); // use _zero reg to clear memory (shorter code) 1.179 + if (UseIncDec) { 1.180 + shrl(index, 3); // divide by 8 and set carry flag if bit 2 was set 1.181 + } else { 1.182 + shrl(index, 2); // use 2 instructions to avoid partial flag stall 1.183 + shrl(index, 1); 1.184 + } 1.185 + // index could have been not a multiple of 8 (i.e., bit 2 was set) 1.186 + { Label even; 1.187 + // note: if index was a multiple of 8, than it cannot 1.188 + // be 0 now otherwise it must have been 0 before 1.189 + // => if it is even, we don't need to check for 0 again 1.190 + jcc(Assembler::carryClear, even); 1.191 + // clear topmost word (no jump needed if conditional assignment would work here) 1.192 + movl(Address(obj, index, Address::times_8, hdr_size_in_bytes - 0*BytesPerWord), t1); 1.193 + // index could be 0 now, need to check again 1.194 + jcc(Assembler::zero, done); 1.195 + bind(even); 1.196 + } 1.197 + // initialize remaining object fields: rdx is a multiple of 2 now 1.198 + { Label loop; 1.199 + bind(loop); 1.200 + movl(Address(obj, index, Address::times_8, hdr_size_in_bytes - 1*BytesPerWord), t1); 1.201 + movl(Address(obj, index, Address::times_8, hdr_size_in_bytes - 2*BytesPerWord), t1); 1.202 + decrement(index); 1.203 + jcc(Assembler::notZero, loop); 1.204 + } 1.205 + 1.206 + // done 1.207 + bind(done); 1.208 +} 1.209 + 1.210 + 1.211 +void C1_MacroAssembler::allocate_object(Register obj, Register t1, Register t2, int header_size, int object_size, Register klass, Label& slow_case) { 1.212 + assert(obj == rax, "obj must be in rax, for cmpxchg"); 1.213 + assert(obj != t1 && obj != t2 && t1 != t2, "registers must be different"); // XXX really? 1.214 + assert(header_size >= 0 && object_size >= header_size, "illegal sizes"); 1.215 + 1.216 + try_allocate(obj, noreg, object_size * BytesPerWord, t1, t2, slow_case); 1.217 + 1.218 + initialize_object(obj, klass, noreg, object_size * HeapWordSize, t1, t2); 1.219 +} 1.220 + 1.221 +void C1_MacroAssembler::initialize_object(Register obj, Register klass, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2) { 1.222 + assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0, 1.223 + "con_size_in_bytes is not multiple of alignment"); 1.224 + const int hdr_size_in_bytes = oopDesc::header_size_in_bytes(); 1.225 + 1.226 + initialize_header(obj, klass, noreg, t1, t2); 1.227 + 1.228 + // clear rest of allocated space 1.229 + const Register t1_zero = t1; 1.230 + const Register index = t2; 1.231 + const int threshold = 6 * BytesPerWord; // approximate break even point for code size (see comments below) 1.232 + if (var_size_in_bytes != noreg) { 1.233 + movl(index, var_size_in_bytes); 1.234 + initialize_body(obj, index, hdr_size_in_bytes, t1_zero); 1.235 + } else if (con_size_in_bytes <= threshold) { 1.236 + // use explicit null stores 1.237 + // code size = 2 + 3*n bytes (n = number of fields to clear) 1.238 + xorl(t1_zero, t1_zero); // use t1_zero reg to clear memory (shorter code) 1.239 + for (int i = hdr_size_in_bytes; i < con_size_in_bytes; i += BytesPerWord) 1.240 + movl(Address(obj, i), t1_zero); 1.241 + } else if (con_size_in_bytes > hdr_size_in_bytes) { 1.242 + // use loop to null out the fields 1.243 + // code size = 16 bytes for even n (n = number of fields to clear) 1.244 + // initialize last object field first if odd number of fields 1.245 + xorl(t1_zero, t1_zero); // use t1_zero reg to clear memory (shorter code) 1.246 + movl(index, (con_size_in_bytes - hdr_size_in_bytes) >> 3); 1.247 + // initialize last object field if constant size is odd 1.248 + if (((con_size_in_bytes - hdr_size_in_bytes) & 4) != 0) 1.249 + movl(Address(obj, con_size_in_bytes - (1*BytesPerWord)), t1_zero); 1.250 + // initialize remaining object fields: rdx is a multiple of 2 1.251 + { Label loop; 1.252 + bind(loop); 1.253 + movl(Address(obj, index, Address::times_8, 1.254 + hdr_size_in_bytes - (1*BytesPerWord)), t1_zero); 1.255 + movl(Address(obj, index, Address::times_8, 1.256 + hdr_size_in_bytes - (2*BytesPerWord)), t1_zero); 1.257 + decrement(index); 1.258 + jcc(Assembler::notZero, loop); 1.259 + } 1.260 + } 1.261 + 1.262 + if (DTraceAllocProbes) { 1.263 + assert(obj == rax, "must be"); 1.264 + call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id))); 1.265 + } 1.266 + 1.267 + verify_oop(obj); 1.268 +} 1.269 + 1.270 +void C1_MacroAssembler::allocate_array(Register obj, Register len, Register t1, Register t2, int header_size, Address::ScaleFactor f, Register klass, Label& slow_case) { 1.271 + assert(obj == rax, "obj must be in rax, for cmpxchg"); 1.272 + assert_different_registers(obj, len, t1, t2, klass); 1.273 + 1.274 + // determine alignment mask 1.275 + assert(BytesPerWord == 4, "must be a multiple of 2 for masking code to work"); 1.276 + 1.277 + // check for negative or excessive length 1.278 + cmpl(len, max_array_allocation_length); 1.279 + jcc(Assembler::above, slow_case); 1.280 + 1.281 + const Register arr_size = t2; // okay to be the same 1.282 + // align object end 1.283 + movl(arr_size, header_size * BytesPerWord + MinObjAlignmentInBytesMask); 1.284 + leal(arr_size, Address(arr_size, len, f)); 1.285 + andl(arr_size, ~MinObjAlignmentInBytesMask); 1.286 + 1.287 + try_allocate(obj, arr_size, 0, t1, t2, slow_case); 1.288 + 1.289 + initialize_header(obj, klass, len, t1, t2); 1.290 + 1.291 + // clear rest of allocated space 1.292 + const Register len_zero = len; 1.293 + initialize_body(obj, arr_size, header_size * BytesPerWord, len_zero); 1.294 + 1.295 + if (DTraceAllocProbes) { 1.296 + assert(obj == rax, "must be"); 1.297 + call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id))); 1.298 + } 1.299 + 1.300 + verify_oop(obj); 1.301 +} 1.302 + 1.303 + 1.304 + 1.305 +void C1_MacroAssembler::inline_cache_check(Register receiver, Register iCache) { 1.306 + verify_oop(receiver); 1.307 + // explicit NULL check not needed since load from [klass_offset] causes a trap 1.308 + // check against inline cache 1.309 + assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()), "must add explicit null check"); 1.310 + int start_offset = offset(); 1.311 + cmpl(iCache, Address(receiver, oopDesc::klass_offset_in_bytes())); 1.312 + // if icache check fails, then jump to runtime routine 1.313 + // Note: RECEIVER must still contain the receiver! 1.314 + jump_cc(Assembler::notEqual, 1.315 + RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 1.316 + assert(offset() - start_offset == 9, "check alignment in emit_method_entry"); 1.317 +} 1.318 + 1.319 + 1.320 +void C1_MacroAssembler::method_exit(bool restore_frame) { 1.321 + if (restore_frame) { 1.322 + leave(); 1.323 + } 1.324 + ret(0); 1.325 +} 1.326 + 1.327 + 1.328 +void C1_MacroAssembler::build_frame(int frame_size_in_bytes) { 1.329 + // Make sure there is enough stack space for this method's activation. 1.330 + // Note that we do this before doing an enter(). This matches the 1.331 + // ordering of C2's stack overflow check / rsp decrement and allows 1.332 + // the SharedRuntime stack overflow handling to be consistent 1.333 + // between the two compilers. 1.334 + generate_stack_overflow_check(frame_size_in_bytes); 1.335 + 1.336 + enter(); 1.337 +#ifdef TIERED 1.338 + // c2 leaves fpu stack dirty. Clean it on entry 1.339 + if (UseSSE < 2 ) { 1.340 + empty_FPU_stack(); 1.341 + } 1.342 +#endif // TIERED 1.343 + decrement(rsp, frame_size_in_bytes); // does not emit code for frame_size == 0 1.344 +} 1.345 + 1.346 + 1.347 +void C1_MacroAssembler::unverified_entry(Register receiver, Register ic_klass) { 1.348 + if (C1Breakpoint) int3(); 1.349 + inline_cache_check(receiver, ic_klass); 1.350 +} 1.351 + 1.352 + 1.353 +void C1_MacroAssembler::verified_entry() { 1.354 + if (C1Breakpoint)int3(); 1.355 + // build frame 1.356 + verify_FPU(0, "method_entry"); 1.357 +} 1.358 + 1.359 + 1.360 +#ifndef PRODUCT 1.361 + 1.362 +void C1_MacroAssembler::verify_stack_oop(int stack_offset) { 1.363 + if (!VerifyOops) return; 1.364 + verify_oop_addr(Address(rsp, stack_offset)); 1.365 +} 1.366 + 1.367 +void C1_MacroAssembler::verify_not_null_oop(Register r) { 1.368 + if (!VerifyOops) return; 1.369 + Label not_null; 1.370 + testl(r, r); 1.371 + jcc(Assembler::notZero, not_null); 1.372 + stop("non-null oop required"); 1.373 + bind(not_null); 1.374 + verify_oop(r); 1.375 +} 1.376 + 1.377 +void C1_MacroAssembler::invalidate_registers(bool inv_rax, bool inv_rbx, bool inv_rcx, bool inv_rdx, bool inv_rsi, bool inv_rdi) { 1.378 +#ifdef ASSERT 1.379 + if (inv_rax) movl(rax, 0xDEAD); 1.380 + if (inv_rbx) movl(rbx, 0xDEAD); 1.381 + if (inv_rcx) movl(rcx, 0xDEAD); 1.382 + if (inv_rdx) movl(rdx, 0xDEAD); 1.383 + if (inv_rsi) movl(rsi, 0xDEAD); 1.384 + if (inv_rdi) movl(rdi, 0xDEAD); 1.385 +#endif 1.386 +} 1.387 + 1.388 +#endif // ifndef PRODUCT