1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/src/cpu/x86/vm/c1_LinearScan_x86.cpp Wed Apr 27 01:25:04 2016 +0800
1.3 @@ -0,0 +1,1242 @@
1.4 +/*
1.5 + * Copyright (c) 2005, 2014, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
1.23 + * or visit www.oracle.com if you need additional information or have any
1.24 + * questions.
1.25 + *
1.26 + */
1.27 +
1.28 +#include "precompiled.hpp"
1.29 +#include "c1/c1_Instruction.hpp"
1.30 +#include "c1/c1_LinearScan.hpp"
1.31 +#include "utilities/bitMap.inline.hpp"
1.32 +
1.33 +
1.34 +//----------------------------------------------------------------------
1.35 +// Allocation of FPU stack slots (Intel x86 only)
1.36 +//----------------------------------------------------------------------
1.37 +
1.38 +void LinearScan::allocate_fpu_stack() {
1.39 + // First compute which FPU registers are live at the start of each basic block
1.40 + // (To minimize the amount of work we have to do if we have to merge FPU stacks)
1.41 + if (ComputeExactFPURegisterUsage) {
1.42 + Interval* intervals_in_register, *intervals_in_memory;
1.43 + create_unhandled_lists(&intervals_in_register, &intervals_in_memory, is_in_fpu_register, NULL);
1.44 +
1.45 + // ignore memory intervals by overwriting intervals_in_memory
1.46 + // the dummy interval is needed to enforce the walker to walk until the given id:
1.47 + // without it, the walker stops when the unhandled-list is empty -> live information
1.48 + // beyond this point would be incorrect.
1.49 + Interval* dummy_interval = new Interval(any_reg);
1.50 + dummy_interval->add_range(max_jint - 2, max_jint - 1);
1.51 + dummy_interval->set_next(Interval::end());
1.52 + intervals_in_memory = dummy_interval;
1.53 +
1.54 + IntervalWalker iw(this, intervals_in_register, intervals_in_memory);
1.55 +
1.56 + const int num_blocks = block_count();
1.57 + for (int i = 0; i < num_blocks; i++) {
1.58 + BlockBegin* b = block_at(i);
1.59 +
1.60 + // register usage is only needed for merging stacks -> compute only
1.61 + // when more than one predecessor.
1.62 + // the block must not have any spill moves at the beginning (checked by assertions)
1.63 + // spill moves would use intervals that are marked as handled and so the usage bit
1.64 + // would been set incorrectly
1.65 +
1.66 + // NOTE: the check for number_of_preds > 1 is necessary. A block with only one
1.67 + // predecessor may have spill moves at the begin of the block.
1.68 + // If an interval ends at the current instruction id, it is not possible
1.69 + // to decide if the register is live or not at the block begin -> the
1.70 + // register information would be incorrect.
1.71 + if (b->number_of_preds() > 1) {
1.72 + int id = b->first_lir_instruction_id();
1.73 + BitMap regs(FrameMap::nof_fpu_regs);
1.74 + regs.clear();
1.75 +
1.76 + iw.walk_to(id); // walk after the first instruction (always a label) of the block
1.77 + assert(iw.current_position() == id, "did not walk completely to id");
1.78 +
1.79 + // Only consider FPU values in registers
1.80 + Interval* interval = iw.active_first(fixedKind);
1.81 + while (interval != Interval::end()) {
1.82 + int reg = interval->assigned_reg();
1.83 + assert(reg >= pd_first_fpu_reg && reg <= pd_last_fpu_reg, "no fpu register");
1.84 + assert(interval->assigned_regHi() == -1, "must not have hi register (doubles stored in one register)");
1.85 + assert(interval->from() <= id && id < interval->to(), "interval out of range");
1.86 +
1.87 +#ifndef PRODUCT
1.88 + if (TraceFPURegisterUsage) {
1.89 + tty->print("fpu reg %d is live because of ", reg - pd_first_fpu_reg); interval->print();
1.90 + }
1.91 +#endif
1.92 +
1.93 + regs.set_bit(reg - pd_first_fpu_reg);
1.94 + interval = interval->next();
1.95 + }
1.96 +
1.97 + b->set_fpu_register_usage(regs);
1.98 +
1.99 +#ifndef PRODUCT
1.100 + if (TraceFPURegisterUsage) {
1.101 + tty->print("FPU regs for block %d, LIR instr %d): ", b->block_id(), id); regs.print_on(tty); tty->cr();
1.102 + }
1.103 +#endif
1.104 + }
1.105 + }
1.106 + }
1.107 +
1.108 + FpuStackAllocator alloc(ir()->compilation(), this);
1.109 + _fpu_stack_allocator = &alloc;
1.110 + alloc.allocate();
1.111 + _fpu_stack_allocator = NULL;
1.112 +}
1.113 +
1.114 +
1.115 +FpuStackAllocator::FpuStackAllocator(Compilation* compilation, LinearScan* allocator)
1.116 + : _compilation(compilation)
1.117 + , _lir(NULL)
1.118 + , _pos(-1)
1.119 + , _allocator(allocator)
1.120 + , _sim(compilation)
1.121 + , _temp_sim(compilation)
1.122 +{}
1.123 +
1.124 +void FpuStackAllocator::allocate() {
1.125 + int num_blocks = allocator()->block_count();
1.126 + for (int i = 0; i < num_blocks; i++) {
1.127 + // Set up to process block
1.128 + BlockBegin* block = allocator()->block_at(i);
1.129 + intArray* fpu_stack_state = block->fpu_stack_state();
1.130 +
1.131 +#ifndef PRODUCT
1.132 + if (TraceFPUStack) {
1.133 + tty->cr();
1.134 + tty->print_cr("------- Begin of new Block %d -------", block->block_id());
1.135 + }
1.136 +#endif
1.137 +
1.138 + assert(fpu_stack_state != NULL ||
1.139 + block->end()->as_Base() != NULL ||
1.140 + block->is_set(BlockBegin::exception_entry_flag),
1.141 + "FPU stack state must be present due to linear-scan order for FPU stack allocation");
1.142 + // note: exception handler entries always start with an empty fpu stack
1.143 + // because stack merging would be too complicated
1.144 +
1.145 + if (fpu_stack_state != NULL) {
1.146 + sim()->read_state(fpu_stack_state);
1.147 + } else {
1.148 + sim()->clear();
1.149 + }
1.150 +
1.151 +#ifndef PRODUCT
1.152 + if (TraceFPUStack) {
1.153 + tty->print("Reading FPU state for block %d:", block->block_id());
1.154 + sim()->print();
1.155 + tty->cr();
1.156 + }
1.157 +#endif
1.158 +
1.159 + allocate_block(block);
1.160 + CHECK_BAILOUT();
1.161 + }
1.162 +}
1.163 +
1.164 +void FpuStackAllocator::allocate_block(BlockBegin* block) {
1.165 + bool processed_merge = false;
1.166 + LIR_OpList* insts = block->lir()->instructions_list();
1.167 + set_lir(block->lir());
1.168 + set_pos(0);
1.169 +
1.170 +
1.171 + // Note: insts->length() may change during loop
1.172 + while (pos() < insts->length()) {
1.173 + LIR_Op* op = insts->at(pos());
1.174 + _debug_information_computed = false;
1.175 +
1.176 +#ifndef PRODUCT
1.177 + if (TraceFPUStack) {
1.178 + op->print();
1.179 + }
1.180 + check_invalid_lir_op(op);
1.181 +#endif
1.182 +
1.183 + LIR_OpBranch* branch = op->as_OpBranch();
1.184 + LIR_Op1* op1 = op->as_Op1();
1.185 + LIR_Op2* op2 = op->as_Op2();
1.186 + LIR_OpCall* opCall = op->as_OpCall();
1.187 +
1.188 + if (branch != NULL && branch->block() != NULL) {
1.189 + if (!processed_merge) {
1.190 + // propagate stack at first branch to a successor
1.191 + processed_merge = true;
1.192 + bool required_merge = merge_fpu_stack_with_successors(block);
1.193 +
1.194 + assert(!required_merge || branch->cond() == lir_cond_always, "splitting of critical edges should prevent FPU stack mismatches at cond branches");
1.195 + }
1.196 +
1.197 + } else if (op1 != NULL) {
1.198 + handle_op1(op1);
1.199 + } else if (op2 != NULL) {
1.200 + handle_op2(op2);
1.201 + } else if (opCall != NULL) {
1.202 + handle_opCall(opCall);
1.203 + }
1.204 +
1.205 + compute_debug_information(op);
1.206 +
1.207 + set_pos(1 + pos());
1.208 + }
1.209 +
1.210 + // Propagate stack when block does not end with branch
1.211 + if (!processed_merge) {
1.212 + merge_fpu_stack_with_successors(block);
1.213 + }
1.214 +}
1.215 +
1.216 +
1.217 +void FpuStackAllocator::compute_debug_information(LIR_Op* op) {
1.218 + if (!_debug_information_computed && op->id() != -1 && allocator()->has_info(op->id())) {
1.219 + visitor.visit(op);
1.220 +
1.221 + // exception handling
1.222 + if (allocator()->compilation()->has_exception_handlers()) {
1.223 + XHandlers* xhandlers = visitor.all_xhandler();
1.224 + int n = xhandlers->length();
1.225 + for (int k = 0; k < n; k++) {
1.226 + allocate_exception_handler(xhandlers->handler_at(k));
1.227 + }
1.228 + } else {
1.229 + assert(visitor.all_xhandler()->length() == 0, "missed exception handler");
1.230 + }
1.231 +
1.232 + // compute debug information
1.233 + int n = visitor.info_count();
1.234 + assert(n > 0, "should not visit operation otherwise");
1.235 +
1.236 + for (int j = 0; j < n; j++) {
1.237 + CodeEmitInfo* info = visitor.info_at(j);
1.238 + // Compute debug information
1.239 + allocator()->compute_debug_info(info, op->id());
1.240 + }
1.241 + }
1.242 + _debug_information_computed = true;
1.243 +}
1.244 +
1.245 +void FpuStackAllocator::allocate_exception_handler(XHandler* xhandler) {
1.246 + if (!sim()->is_empty()) {
1.247 + LIR_List* old_lir = lir();
1.248 + int old_pos = pos();
1.249 + intArray* old_state = sim()->write_state();
1.250 +
1.251 +#ifndef PRODUCT
1.252 + if (TraceFPUStack) {
1.253 + tty->cr();
1.254 + tty->print_cr("------- begin of exception handler -------");
1.255 + }
1.256 +#endif
1.257 +
1.258 + if (xhandler->entry_code() == NULL) {
1.259 + // need entry code to clear FPU stack
1.260 + LIR_List* entry_code = new LIR_List(_compilation);
1.261 + entry_code->jump(xhandler->entry_block());
1.262 + xhandler->set_entry_code(entry_code);
1.263 + }
1.264 +
1.265 + LIR_OpList* insts = xhandler->entry_code()->instructions_list();
1.266 + set_lir(xhandler->entry_code());
1.267 + set_pos(0);
1.268 +
1.269 + // Note: insts->length() may change during loop
1.270 + while (pos() < insts->length()) {
1.271 + LIR_Op* op = insts->at(pos());
1.272 +
1.273 +#ifndef PRODUCT
1.274 + if (TraceFPUStack) {
1.275 + op->print();
1.276 + }
1.277 + check_invalid_lir_op(op);
1.278 +#endif
1.279 +
1.280 + switch (op->code()) {
1.281 + case lir_move:
1.282 + assert(op->as_Op1() != NULL, "must be LIR_Op1");
1.283 + assert(pos() != insts->length() - 1, "must not be last operation");
1.284 +
1.285 + handle_op1((LIR_Op1*)op);
1.286 + break;
1.287 +
1.288 + case lir_branch:
1.289 + assert(op->as_OpBranch()->cond() == lir_cond_always, "must be unconditional branch");
1.290 + assert(pos() == insts->length() - 1, "must be last operation");
1.291 +
1.292 + // remove all remaining dead registers from FPU stack
1.293 + clear_fpu_stack(LIR_OprFact::illegalOpr);
1.294 + break;
1.295 +
1.296 + default:
1.297 + // other operations not allowed in exception entry code
1.298 + ShouldNotReachHere();
1.299 + }
1.300 +
1.301 + set_pos(pos() + 1);
1.302 + }
1.303 +
1.304 +#ifndef PRODUCT
1.305 + if (TraceFPUStack) {
1.306 + tty->cr();
1.307 + tty->print_cr("------- end of exception handler -------");
1.308 + }
1.309 +#endif
1.310 +
1.311 + set_lir(old_lir);
1.312 + set_pos(old_pos);
1.313 + sim()->read_state(old_state);
1.314 + }
1.315 +}
1.316 +
1.317 +
1.318 +int FpuStackAllocator::fpu_num(LIR_Opr opr) {
1.319 + assert(opr->is_fpu_register() && !opr->is_xmm_register(), "shouldn't call this otherwise");
1.320 + return opr->is_single_fpu() ? opr->fpu_regnr() : opr->fpu_regnrLo();
1.321 +}
1.322 +
1.323 +int FpuStackAllocator::tos_offset(LIR_Opr opr) {
1.324 + return sim()->offset_from_tos(fpu_num(opr));
1.325 +}
1.326 +
1.327 +
1.328 +LIR_Opr FpuStackAllocator::to_fpu_stack(LIR_Opr opr) {
1.329 + assert(opr->is_fpu_register() && !opr->is_xmm_register(), "shouldn't call this otherwise");
1.330 +
1.331 + int stack_offset = tos_offset(opr);
1.332 + if (opr->is_single_fpu()) {
1.333 + return LIR_OprFact::single_fpu(stack_offset)->make_fpu_stack_offset();
1.334 + } else {
1.335 + assert(opr->is_double_fpu(), "shouldn't call this otherwise");
1.336 + return LIR_OprFact::double_fpu(stack_offset)->make_fpu_stack_offset();
1.337 + }
1.338 +}
1.339 +
1.340 +LIR_Opr FpuStackAllocator::to_fpu_stack_top(LIR_Opr opr, bool dont_check_offset) {
1.341 + assert(opr->is_fpu_register() && !opr->is_xmm_register(), "shouldn't call this otherwise");
1.342 + assert(dont_check_offset || tos_offset(opr) == 0, "operand is not on stack top");
1.343 +
1.344 + int stack_offset = 0;
1.345 + if (opr->is_single_fpu()) {
1.346 + return LIR_OprFact::single_fpu(stack_offset)->make_fpu_stack_offset();
1.347 + } else {
1.348 + assert(opr->is_double_fpu(), "shouldn't call this otherwise");
1.349 + return LIR_OprFact::double_fpu(stack_offset)->make_fpu_stack_offset();
1.350 + }
1.351 +}
1.352 +
1.353 +
1.354 +
1.355 +void FpuStackAllocator::insert_op(LIR_Op* op) {
1.356 + lir()->insert_before(pos(), op);
1.357 + set_pos(1 + pos());
1.358 +}
1.359 +
1.360 +
1.361 +void FpuStackAllocator::insert_exchange(int offset) {
1.362 + if (offset > 0) {
1.363 + LIR_Op1* fxch_op = new LIR_Op1(lir_fxch, LIR_OprFact::intConst(offset), LIR_OprFact::illegalOpr);
1.364 + insert_op(fxch_op);
1.365 + sim()->swap(offset);
1.366 +
1.367 +#ifndef PRODUCT
1.368 + if (TraceFPUStack) {
1.369 + tty->print("Exchanged register: %d New state: ", sim()->get_slot(0)); sim()->print(); tty->cr();
1.370 + }
1.371 +#endif
1.372 +
1.373 + }
1.374 +}
1.375 +
1.376 +void FpuStackAllocator::insert_exchange(LIR_Opr opr) {
1.377 + insert_exchange(tos_offset(opr));
1.378 +}
1.379 +
1.380 +
1.381 +void FpuStackAllocator::insert_free(int offset) {
1.382 + // move stack slot to the top of stack and then pop it
1.383 + insert_exchange(offset);
1.384 +
1.385 + LIR_Op* fpop = new LIR_Op0(lir_fpop_raw);
1.386 + insert_op(fpop);
1.387 + sim()->pop();
1.388 +
1.389 +#ifndef PRODUCT
1.390 + if (TraceFPUStack) {
1.391 + tty->print("Inserted pop New state: "); sim()->print(); tty->cr();
1.392 + }
1.393 +#endif
1.394 +}
1.395 +
1.396 +
1.397 +void FpuStackAllocator::insert_free_if_dead(LIR_Opr opr) {
1.398 + if (sim()->contains(fpu_num(opr))) {
1.399 + int res_slot = tos_offset(opr);
1.400 + insert_free(res_slot);
1.401 + }
1.402 +}
1.403 +
1.404 +void FpuStackAllocator::insert_free_if_dead(LIR_Opr opr, LIR_Opr ignore) {
1.405 + if (fpu_num(opr) != fpu_num(ignore) && sim()->contains(fpu_num(opr))) {
1.406 + int res_slot = tos_offset(opr);
1.407 + insert_free(res_slot);
1.408 + }
1.409 +}
1.410 +
1.411 +void FpuStackAllocator::insert_copy(LIR_Opr from, LIR_Opr to) {
1.412 + int offset = tos_offset(from);
1.413 + LIR_Op1* fld = new LIR_Op1(lir_fld, LIR_OprFact::intConst(offset), LIR_OprFact::illegalOpr);
1.414 + insert_op(fld);
1.415 +
1.416 + sim()->push(fpu_num(to));
1.417 +
1.418 +#ifndef PRODUCT
1.419 + if (TraceFPUStack) {
1.420 + tty->print("Inserted copy (%d -> %d) New state: ", fpu_num(from), fpu_num(to)); sim()->print(); tty->cr();
1.421 + }
1.422 +#endif
1.423 +}
1.424 +
1.425 +void FpuStackAllocator::do_rename(LIR_Opr from, LIR_Opr to) {
1.426 + sim()->rename(fpu_num(from), fpu_num(to));
1.427 +}
1.428 +
1.429 +void FpuStackAllocator::do_push(LIR_Opr opr) {
1.430 + sim()->push(fpu_num(opr));
1.431 +}
1.432 +
1.433 +void FpuStackAllocator::pop_if_last_use(LIR_Op* op, LIR_Opr opr) {
1.434 + assert(op->fpu_pop_count() == 0, "fpu_pop_count alredy set");
1.435 + assert(tos_offset(opr) == 0, "can only pop stack top");
1.436 +
1.437 + if (opr->is_last_use()) {
1.438 + op->set_fpu_pop_count(1);
1.439 + sim()->pop();
1.440 + }
1.441 +}
1.442 +
1.443 +void FpuStackAllocator::pop_always(LIR_Op* op, LIR_Opr opr) {
1.444 + assert(op->fpu_pop_count() == 0, "fpu_pop_count alredy set");
1.445 + assert(tos_offset(opr) == 0, "can only pop stack top");
1.446 +
1.447 + op->set_fpu_pop_count(1);
1.448 + sim()->pop();
1.449 +}
1.450 +
1.451 +void FpuStackAllocator::clear_fpu_stack(LIR_Opr preserve) {
1.452 + int result_stack_size = (preserve->is_fpu_register() && !preserve->is_xmm_register() ? 1 : 0);
1.453 + while (sim()->stack_size() > result_stack_size) {
1.454 + assert(!sim()->slot_is_empty(0), "not allowed");
1.455 +
1.456 + if (result_stack_size == 0 || sim()->get_slot(0) != fpu_num(preserve)) {
1.457 + insert_free(0);
1.458 + } else {
1.459 + // move "preserve" to bottom of stack so that all other stack slots can be popped
1.460 + insert_exchange(sim()->stack_size() - 1);
1.461 + }
1.462 + }
1.463 +}
1.464 +
1.465 +
1.466 +void FpuStackAllocator::handle_op1(LIR_Op1* op1) {
1.467 + LIR_Opr in = op1->in_opr();
1.468 + LIR_Opr res = op1->result_opr();
1.469 +
1.470 + LIR_Opr new_in = in; // new operands relative to the actual fpu stack top
1.471 + LIR_Opr new_res = res;
1.472 +
1.473 + // Note: this switch is processed for all LIR_Op1, regardless if they have FPU-arguments,
1.474 + // so checks for is_float_kind() are necessary inside the cases
1.475 + switch (op1->code()) {
1.476 +
1.477 + case lir_return: {
1.478 + // FPU-Stack must only contain the (optional) fpu return value.
1.479 + // All remaining dead values are popped from the stack
1.480 + // If the input operand is a fpu-register, it is exchanged to the bottom of the stack
1.481 +
1.482 + clear_fpu_stack(in);
1.483 + if (in->is_fpu_register() && !in->is_xmm_register()) {
1.484 + new_in = to_fpu_stack_top(in);
1.485 + }
1.486 +
1.487 + break;
1.488 + }
1.489 +
1.490 + case lir_move: {
1.491 + if (in->is_fpu_register() && !in->is_xmm_register()) {
1.492 + if (res->is_xmm_register()) {
1.493 + // move from fpu register to xmm register (necessary for operations that
1.494 + // are not available in the SSE instruction set)
1.495 + insert_exchange(in);
1.496 + new_in = to_fpu_stack_top(in);
1.497 + pop_always(op1, in);
1.498 +
1.499 + } else if (res->is_fpu_register() && !res->is_xmm_register()) {
1.500 + // move from fpu-register to fpu-register:
1.501 + // * input and result register equal:
1.502 + // nothing to do
1.503 + // * input register is last use:
1.504 + // rename the input register to result register -> input register
1.505 + // not present on fpu-stack afterwards
1.506 + // * input register not last use:
1.507 + // duplicate input register to result register to preserve input
1.508 + //
1.509 + // Note: The LIR-Assembler does not produce any code for fpu register moves,
1.510 + // so input and result stack index must be equal
1.511 +
1.512 + if (fpu_num(in) == fpu_num(res)) {
1.513 + // nothing to do
1.514 + } else if (in->is_last_use()) {
1.515 + insert_free_if_dead(res);//, in);
1.516 + do_rename(in, res);
1.517 + } else {
1.518 + insert_free_if_dead(res);
1.519 + insert_copy(in, res);
1.520 + }
1.521 + new_in = to_fpu_stack(res);
1.522 + new_res = new_in;
1.523 +
1.524 + } else {
1.525 + // move from fpu-register to memory
1.526 + // input operand must be on top of stack
1.527 +
1.528 + insert_exchange(in);
1.529 +
1.530 + // create debug information here because afterwards the register may have been popped
1.531 + compute_debug_information(op1);
1.532 +
1.533 + new_in = to_fpu_stack_top(in);
1.534 + pop_if_last_use(op1, in);
1.535 + }
1.536 +
1.537 + } else if (res->is_fpu_register() && !res->is_xmm_register()) {
1.538 + // move from memory/constant to fpu register
1.539 + // result is pushed on the stack
1.540 +
1.541 + insert_free_if_dead(res);
1.542 +
1.543 + // create debug information before register is pushed
1.544 + compute_debug_information(op1);
1.545 +
1.546 + do_push(res);
1.547 + new_res = to_fpu_stack_top(res);
1.548 + }
1.549 + break;
1.550 + }
1.551 +
1.552 + case lir_neg: {
1.553 + if (in->is_fpu_register() && !in->is_xmm_register()) {
1.554 + assert(res->is_fpu_register() && !res->is_xmm_register(), "must be");
1.555 + assert(in->is_last_use(), "old value gets destroyed");
1.556 +
1.557 + insert_free_if_dead(res, in);
1.558 + insert_exchange(in);
1.559 + new_in = to_fpu_stack_top(in);
1.560 +
1.561 + do_rename(in, res);
1.562 + new_res = to_fpu_stack_top(res);
1.563 + }
1.564 + break;
1.565 + }
1.566 +
1.567 + case lir_convert: {
1.568 + Bytecodes::Code bc = op1->as_OpConvert()->bytecode();
1.569 + switch (bc) {
1.570 + case Bytecodes::_d2f:
1.571 + case Bytecodes::_f2d:
1.572 + assert(res->is_fpu_register(), "must be");
1.573 + assert(in->is_fpu_register(), "must be");
1.574 +
1.575 + if (!in->is_xmm_register() && !res->is_xmm_register()) {
1.576 + // this is quite the same as a move from fpu-register to fpu-register
1.577 + // Note: input and result operands must have different types
1.578 + if (fpu_num(in) == fpu_num(res)) {
1.579 + // nothing to do
1.580 + new_in = to_fpu_stack(in);
1.581 + } else if (in->is_last_use()) {
1.582 + insert_free_if_dead(res);//, in);
1.583 + new_in = to_fpu_stack(in);
1.584 + do_rename(in, res);
1.585 + } else {
1.586 + insert_free_if_dead(res);
1.587 + insert_copy(in, res);
1.588 + new_in = to_fpu_stack_top(in, true);
1.589 + }
1.590 + new_res = to_fpu_stack(res);
1.591 + }
1.592 +
1.593 + break;
1.594 +
1.595 + case Bytecodes::_i2f:
1.596 + case Bytecodes::_l2f:
1.597 + case Bytecodes::_i2d:
1.598 + case Bytecodes::_l2d:
1.599 + assert(res->is_fpu_register(), "must be");
1.600 + if (!res->is_xmm_register()) {
1.601 + insert_free_if_dead(res);
1.602 + do_push(res);
1.603 + new_res = to_fpu_stack_top(res);
1.604 + }
1.605 + break;
1.606 +
1.607 + case Bytecodes::_f2i:
1.608 + case Bytecodes::_d2i:
1.609 + assert(in->is_fpu_register(), "must be");
1.610 + if (!in->is_xmm_register()) {
1.611 + insert_exchange(in);
1.612 + new_in = to_fpu_stack_top(in);
1.613 +
1.614 + // TODO: update registes of stub
1.615 + }
1.616 + break;
1.617 +
1.618 + case Bytecodes::_f2l:
1.619 + case Bytecodes::_d2l:
1.620 + assert(in->is_fpu_register(), "must be");
1.621 + if (!in->is_xmm_register()) {
1.622 + insert_exchange(in);
1.623 + new_in = to_fpu_stack_top(in);
1.624 + pop_always(op1, in);
1.625 + }
1.626 + break;
1.627 +
1.628 + case Bytecodes::_i2l:
1.629 + case Bytecodes::_l2i:
1.630 + case Bytecodes::_i2b:
1.631 + case Bytecodes::_i2c:
1.632 + case Bytecodes::_i2s:
1.633 + // no fpu operands
1.634 + break;
1.635 +
1.636 + default:
1.637 + ShouldNotReachHere();
1.638 + }
1.639 + break;
1.640 + }
1.641 +
1.642 + case lir_roundfp: {
1.643 + assert(in->is_fpu_register() && !in->is_xmm_register(), "input must be in register");
1.644 + assert(res->is_stack(), "result must be on stack");
1.645 +
1.646 + insert_exchange(in);
1.647 + new_in = to_fpu_stack_top(in);
1.648 + pop_if_last_use(op1, in);
1.649 + break;
1.650 + }
1.651 +
1.652 + default: {
1.653 + assert(!in->is_float_kind() && !res->is_float_kind(), "missed a fpu-operation");
1.654 + }
1.655 + }
1.656 +
1.657 + op1->set_in_opr(new_in);
1.658 + op1->set_result_opr(new_res);
1.659 +}
1.660 +
1.661 +void FpuStackAllocator::handle_op2(LIR_Op2* op2) {
1.662 + LIR_Opr left = op2->in_opr1();
1.663 + if (!left->is_float_kind()) {
1.664 + return;
1.665 + }
1.666 + if (left->is_xmm_register()) {
1.667 + return;
1.668 + }
1.669 +
1.670 + LIR_Opr right = op2->in_opr2();
1.671 + LIR_Opr res = op2->result_opr();
1.672 + LIR_Opr new_left = left; // new operands relative to the actual fpu stack top
1.673 + LIR_Opr new_right = right;
1.674 + LIR_Opr new_res = res;
1.675 +
1.676 + assert(!left->is_xmm_register() && !right->is_xmm_register() && !res->is_xmm_register(), "not for xmm registers");
1.677 +
1.678 + switch (op2->code()) {
1.679 + case lir_cmp:
1.680 + case lir_cmp_fd2i:
1.681 + case lir_ucmp_fd2i:
1.682 + case lir_assert: {
1.683 + assert(left->is_fpu_register(), "invalid LIR");
1.684 + assert(right->is_fpu_register(), "invalid LIR");
1.685 +
1.686 + // the left-hand side must be on top of stack.
1.687 + // the right-hand side is never popped, even if is_last_use is set
1.688 + insert_exchange(left);
1.689 + new_left = to_fpu_stack_top(left);
1.690 + new_right = to_fpu_stack(right);
1.691 + pop_if_last_use(op2, left);
1.692 + break;
1.693 + }
1.694 +
1.695 + case lir_mul_strictfp:
1.696 + case lir_div_strictfp: {
1.697 + assert(op2->tmp1_opr()->is_fpu_register(), "strict operations need temporary fpu stack slot");
1.698 + insert_free_if_dead(op2->tmp1_opr());
1.699 + assert(sim()->stack_size() <= 7, "at least one stack slot must be free");
1.700 + // fall-through: continue with the normal handling of lir_mul and lir_div
1.701 + }
1.702 + case lir_add:
1.703 + case lir_sub:
1.704 + case lir_mul:
1.705 + case lir_div: {
1.706 + assert(left->is_fpu_register(), "must be");
1.707 + assert(res->is_fpu_register(), "must be");
1.708 + assert(left->is_equal(res), "must be");
1.709 +
1.710 + // either the left-hand or the right-hand side must be on top of stack
1.711 + // (if right is not a register, left must be on top)
1.712 + if (!right->is_fpu_register()) {
1.713 + insert_exchange(left);
1.714 + new_left = to_fpu_stack_top(left);
1.715 + } else {
1.716 + // no exchange necessary if right is alredy on top of stack
1.717 + if (tos_offset(right) == 0) {
1.718 + new_left = to_fpu_stack(left);
1.719 + new_right = to_fpu_stack_top(right);
1.720 + } else {
1.721 + insert_exchange(left);
1.722 + new_left = to_fpu_stack_top(left);
1.723 + new_right = to_fpu_stack(right);
1.724 + }
1.725 +
1.726 + if (right->is_last_use()) {
1.727 + op2->set_fpu_pop_count(1);
1.728 +
1.729 + if (tos_offset(right) == 0) {
1.730 + sim()->pop();
1.731 + } else {
1.732 + // if left is on top of stack, the result is placed in the stack
1.733 + // slot of right, so a renaming from right to res is necessary
1.734 + assert(tos_offset(left) == 0, "must be");
1.735 + sim()->pop();
1.736 + do_rename(right, res);
1.737 + }
1.738 + }
1.739 + }
1.740 + new_res = to_fpu_stack(res);
1.741 +
1.742 + break;
1.743 + }
1.744 +
1.745 + case lir_rem: {
1.746 + assert(left->is_fpu_register(), "must be");
1.747 + assert(right->is_fpu_register(), "must be");
1.748 + assert(res->is_fpu_register(), "must be");
1.749 + assert(left->is_equal(res), "must be");
1.750 +
1.751 + // Must bring both operands to top of stack with following operand ordering:
1.752 + // * fpu stack before rem: ... right left
1.753 + // * fpu stack after rem: ... left
1.754 + if (tos_offset(right) != 1) {
1.755 + insert_exchange(right);
1.756 + insert_exchange(1);
1.757 + }
1.758 + insert_exchange(left);
1.759 + assert(tos_offset(right) == 1, "check");
1.760 + assert(tos_offset(left) == 0, "check");
1.761 +
1.762 + new_left = to_fpu_stack_top(left);
1.763 + new_right = to_fpu_stack(right);
1.764 +
1.765 + op2->set_fpu_pop_count(1);
1.766 + sim()->pop();
1.767 + do_rename(right, res);
1.768 +
1.769 + new_res = to_fpu_stack_top(res);
1.770 + break;
1.771 + }
1.772 +
1.773 + case lir_abs:
1.774 + case lir_sqrt: {
1.775 + // Right argument appears to be unused
1.776 + assert(right->is_illegal(), "must be");
1.777 + assert(left->is_fpu_register(), "must be");
1.778 + assert(res->is_fpu_register(), "must be");
1.779 + assert(left->is_last_use(), "old value gets destroyed");
1.780 +
1.781 + insert_free_if_dead(res, left);
1.782 + insert_exchange(left);
1.783 + do_rename(left, res);
1.784 +
1.785 + new_left = to_fpu_stack_top(res);
1.786 + new_res = new_left;
1.787 +
1.788 + op2->set_fpu_stack_size(sim()->stack_size());
1.789 + break;
1.790 + }
1.791 +
1.792 + case lir_log:
1.793 + case lir_log10: {
1.794 + // log and log10 need one temporary fpu stack slot, so
1.795 + // there is one temporary registers stored in temp of the
1.796 + // operation. the stack allocator must guarantee that the stack
1.797 + // slots are really free, otherwise there might be a stack
1.798 + // overflow.
1.799 + assert(right->is_illegal(), "must be");
1.800 + assert(left->is_fpu_register(), "must be");
1.801 + assert(res->is_fpu_register(), "must be");
1.802 + assert(op2->tmp1_opr()->is_fpu_register(), "must be");
1.803 +
1.804 + insert_free_if_dead(op2->tmp1_opr());
1.805 + insert_free_if_dead(res, left);
1.806 + insert_exchange(left);
1.807 + do_rename(left, res);
1.808 +
1.809 + new_left = to_fpu_stack_top(res);
1.810 + new_res = new_left;
1.811 +
1.812 + op2->set_fpu_stack_size(sim()->stack_size());
1.813 + assert(sim()->stack_size() <= 7, "at least one stack slot must be free");
1.814 + break;
1.815 + }
1.816 +
1.817 +
1.818 + case lir_tan:
1.819 + case lir_sin:
1.820 + case lir_cos:
1.821 + case lir_exp: {
1.822 + // sin, cos and exp need two temporary fpu stack slots, so there are two temporary
1.823 + // registers (stored in right and temp of the operation).
1.824 + // the stack allocator must guarantee that the stack slots are really free,
1.825 + // otherwise there might be a stack overflow.
1.826 + assert(left->is_fpu_register(), "must be");
1.827 + assert(res->is_fpu_register(), "must be");
1.828 + // assert(left->is_last_use(), "old value gets destroyed");
1.829 + assert(right->is_fpu_register(), "right is used as the first temporary register");
1.830 + assert(op2->tmp1_opr()->is_fpu_register(), "temp is used as the second temporary register");
1.831 + assert(fpu_num(left) != fpu_num(right) && fpu_num(right) != fpu_num(op2->tmp1_opr()) && fpu_num(op2->tmp1_opr()) != fpu_num(res), "need distinct temp registers");
1.832 +
1.833 + insert_free_if_dead(right);
1.834 + insert_free_if_dead(op2->tmp1_opr());
1.835 +
1.836 + insert_free_if_dead(res, left);
1.837 + insert_exchange(left);
1.838 + do_rename(left, res);
1.839 +
1.840 + new_left = to_fpu_stack_top(res);
1.841 + new_res = new_left;
1.842 +
1.843 + op2->set_fpu_stack_size(sim()->stack_size());
1.844 + assert(sim()->stack_size() <= 6, "at least two stack slots must be free");
1.845 + break;
1.846 + }
1.847 +
1.848 + case lir_pow: {
1.849 + // pow needs two temporary fpu stack slots, so there are two temporary
1.850 + // registers (stored in tmp1 and tmp2 of the operation).
1.851 + // the stack allocator must guarantee that the stack slots are really free,
1.852 + // otherwise there might be a stack overflow.
1.853 + assert(left->is_fpu_register(), "must be");
1.854 + assert(right->is_fpu_register(), "must be");
1.855 + assert(res->is_fpu_register(), "must be");
1.856 +
1.857 + assert(op2->tmp1_opr()->is_fpu_register(), "tmp1 is the first temporary register");
1.858 + assert(op2->tmp2_opr()->is_fpu_register(), "tmp2 is the second temporary register");
1.859 + assert(fpu_num(left) != fpu_num(right) && fpu_num(left) != fpu_num(op2->tmp1_opr()) && fpu_num(left) != fpu_num(op2->tmp2_opr()) && fpu_num(left) != fpu_num(res), "need distinct temp registers");
1.860 + assert(fpu_num(right) != fpu_num(op2->tmp1_opr()) && fpu_num(right) != fpu_num(op2->tmp2_opr()) && fpu_num(right) != fpu_num(res), "need distinct temp registers");
1.861 + assert(fpu_num(op2->tmp1_opr()) != fpu_num(op2->tmp2_opr()) && fpu_num(op2->tmp1_opr()) != fpu_num(res), "need distinct temp registers");
1.862 + assert(fpu_num(op2->tmp2_opr()) != fpu_num(res), "need distinct temp registers");
1.863 +
1.864 + insert_free_if_dead(op2->tmp1_opr());
1.865 + insert_free_if_dead(op2->tmp2_opr());
1.866 +
1.867 + // Must bring both operands to top of stack with following operand ordering:
1.868 + // * fpu stack before pow: ... right left
1.869 + // * fpu stack after pow: ... left
1.870 +
1.871 + insert_free_if_dead(res, right);
1.872 +
1.873 + if (tos_offset(right) != 1) {
1.874 + insert_exchange(right);
1.875 + insert_exchange(1);
1.876 + }
1.877 + insert_exchange(left);
1.878 + assert(tos_offset(right) == 1, "check");
1.879 + assert(tos_offset(left) == 0, "check");
1.880 +
1.881 + new_left = to_fpu_stack_top(left);
1.882 + new_right = to_fpu_stack(right);
1.883 +
1.884 + op2->set_fpu_stack_size(sim()->stack_size());
1.885 + assert(sim()->stack_size() <= 6, "at least two stack slots must be free");
1.886 +
1.887 + sim()->pop();
1.888 +
1.889 + do_rename(right, res);
1.890 +
1.891 + new_res = to_fpu_stack_top(res);
1.892 + break;
1.893 + }
1.894 +
1.895 + default: {
1.896 + assert(false, "missed a fpu-operation");
1.897 + }
1.898 + }
1.899 +
1.900 + op2->set_in_opr1(new_left);
1.901 + op2->set_in_opr2(new_right);
1.902 + op2->set_result_opr(new_res);
1.903 +}
1.904 +
1.905 +void FpuStackAllocator::handle_opCall(LIR_OpCall* opCall) {
1.906 + LIR_Opr res = opCall->result_opr();
1.907 +
1.908 + // clear fpu-stack before call
1.909 + // it may contain dead values that could not have been remved by previous operations
1.910 + clear_fpu_stack(LIR_OprFact::illegalOpr);
1.911 + assert(sim()->is_empty(), "fpu stack must be empty now");
1.912 +
1.913 + // compute debug information before (possible) fpu result is pushed
1.914 + compute_debug_information(opCall);
1.915 +
1.916 + if (res->is_fpu_register() && !res->is_xmm_register()) {
1.917 + do_push(res);
1.918 + opCall->set_result_opr(to_fpu_stack_top(res));
1.919 + }
1.920 +}
1.921 +
1.922 +#ifndef PRODUCT
1.923 +void FpuStackAllocator::check_invalid_lir_op(LIR_Op* op) {
1.924 + switch (op->code()) {
1.925 + case lir_24bit_FPU:
1.926 + case lir_reset_FPU:
1.927 + case lir_ffree:
1.928 + assert(false, "operations not allowed in lir. If one of these operations is needed, check if they have fpu operands");
1.929 + break;
1.930 +
1.931 + case lir_fpop_raw:
1.932 + case lir_fxch:
1.933 + case lir_fld:
1.934 + assert(false, "operations only inserted by FpuStackAllocator");
1.935 + break;
1.936 + }
1.937 +}
1.938 +#endif
1.939 +
1.940 +
1.941 +void FpuStackAllocator::merge_insert_add(LIR_List* instrs, FpuStackSim* cur_sim, int reg) {
1.942 + LIR_Op1* move = new LIR_Op1(lir_move, LIR_OprFact::doubleConst(0), LIR_OprFact::double_fpu(reg)->make_fpu_stack_offset());
1.943 +
1.944 + instrs->instructions_list()->push(move);
1.945 +
1.946 + cur_sim->push(reg);
1.947 + move->set_result_opr(to_fpu_stack(move->result_opr()));
1.948 +
1.949 + #ifndef PRODUCT
1.950 + if (TraceFPUStack) {
1.951 + tty->print("Added new register: %d New state: ", reg); cur_sim->print(); tty->cr();
1.952 + }
1.953 + #endif
1.954 +}
1.955 +
1.956 +void FpuStackAllocator::merge_insert_xchg(LIR_List* instrs, FpuStackSim* cur_sim, int slot) {
1.957 + assert(slot > 0, "no exchange necessary");
1.958 +
1.959 + LIR_Op1* fxch = new LIR_Op1(lir_fxch, LIR_OprFact::intConst(slot));
1.960 + instrs->instructions_list()->push(fxch);
1.961 + cur_sim->swap(slot);
1.962 +
1.963 + #ifndef PRODUCT
1.964 + if (TraceFPUStack) {
1.965 + tty->print("Exchanged register: %d New state: ", cur_sim->get_slot(slot)); cur_sim->print(); tty->cr();
1.966 + }
1.967 + #endif
1.968 +}
1.969 +
1.970 +void FpuStackAllocator::merge_insert_pop(LIR_List* instrs, FpuStackSim* cur_sim) {
1.971 + int reg = cur_sim->get_slot(0);
1.972 +
1.973 + LIR_Op* fpop = new LIR_Op0(lir_fpop_raw);
1.974 + instrs->instructions_list()->push(fpop);
1.975 + cur_sim->pop(reg);
1.976 +
1.977 + #ifndef PRODUCT
1.978 + if (TraceFPUStack) {
1.979 + tty->print("Removed register: %d New state: ", reg); cur_sim->print(); tty->cr();
1.980 + }
1.981 + #endif
1.982 +}
1.983 +
1.984 +bool FpuStackAllocator::merge_rename(FpuStackSim* cur_sim, FpuStackSim* sux_sim, int start_slot, int change_slot) {
1.985 + int reg = cur_sim->get_slot(change_slot);
1.986 +
1.987 + for (int slot = start_slot; slot >= 0; slot--) {
1.988 + int new_reg = sux_sim->get_slot(slot);
1.989 +
1.990 + if (!cur_sim->contains(new_reg)) {
1.991 + cur_sim->set_slot(change_slot, new_reg);
1.992 +
1.993 + #ifndef PRODUCT
1.994 + if (TraceFPUStack) {
1.995 + tty->print("Renamed register %d to %d New state: ", reg, new_reg); cur_sim->print(); tty->cr();
1.996 + }
1.997 + #endif
1.998 +
1.999 + return true;
1.1000 + }
1.1001 + }
1.1002 + return false;
1.1003 +}
1.1004 +
1.1005 +
1.1006 +void FpuStackAllocator::merge_fpu_stack(LIR_List* instrs, FpuStackSim* cur_sim, FpuStackSim* sux_sim) {
1.1007 +#ifndef PRODUCT
1.1008 + if (TraceFPUStack) {
1.1009 + tty->cr();
1.1010 + tty->print("before merging: pred: "); cur_sim->print(); tty->cr();
1.1011 + tty->print(" sux: "); sux_sim->print(); tty->cr();
1.1012 + }
1.1013 +
1.1014 + int slot;
1.1015 + for (slot = 0; slot < cur_sim->stack_size(); slot++) {
1.1016 + assert(!cur_sim->slot_is_empty(slot), "not handled by algorithm");
1.1017 + }
1.1018 + for (slot = 0; slot < sux_sim->stack_size(); slot++) {
1.1019 + assert(!sux_sim->slot_is_empty(slot), "not handled by algorithm");
1.1020 + }
1.1021 +#endif
1.1022 +
1.1023 + // size difference between cur and sux that must be resolved by adding or removing values form the stack
1.1024 + int size_diff = cur_sim->stack_size() - sux_sim->stack_size();
1.1025 +
1.1026 + if (!ComputeExactFPURegisterUsage) {
1.1027 + // add slots that are currently free, but used in successor
1.1028 + // When the exact FPU register usage is computed, the stack does
1.1029 + // not contain dead values at merging -> no values must be added
1.1030 +
1.1031 + int sux_slot = sux_sim->stack_size() - 1;
1.1032 + while (size_diff < 0) {
1.1033 + assert(sux_slot >= 0, "slot out of bounds -> error in algorithm");
1.1034 +
1.1035 + int reg = sux_sim->get_slot(sux_slot);
1.1036 + if (!cur_sim->contains(reg)) {
1.1037 + merge_insert_add(instrs, cur_sim, reg);
1.1038 + size_diff++;
1.1039 +
1.1040 + if (sux_slot + size_diff != 0) {
1.1041 + merge_insert_xchg(instrs, cur_sim, sux_slot + size_diff);
1.1042 + }
1.1043 + }
1.1044 + sux_slot--;
1.1045 + }
1.1046 + }
1.1047 +
1.1048 + assert(cur_sim->stack_size() >= sux_sim->stack_size(), "stack size must be equal or greater now");
1.1049 + assert(size_diff == cur_sim->stack_size() - sux_sim->stack_size(), "must be");
1.1050 +
1.1051 + // stack merge algorithm:
1.1052 + // 1) as long as the current stack top is not in the right location (that meens
1.1053 + // it should not be on the stack top), exchange it into the right location
1.1054 + // 2) if the stack top is right, but the remaining stack is not ordered correctly,
1.1055 + // the stack top is exchanged away to get another value on top ->
1.1056 + // now step 1) can be continued
1.1057 + // the stack can also contain unused items -> these items are removed from stack
1.1058 +
1.1059 + int finished_slot = sux_sim->stack_size() - 1;
1.1060 + while (finished_slot >= 0 || size_diff > 0) {
1.1061 + while (size_diff > 0 || (cur_sim->stack_size() > 0 && cur_sim->get_slot(0) != sux_sim->get_slot(0))) {
1.1062 + int reg = cur_sim->get_slot(0);
1.1063 + if (sux_sim->contains(reg)) {
1.1064 + int sux_slot = sux_sim->offset_from_tos(reg);
1.1065 + merge_insert_xchg(instrs, cur_sim, sux_slot + size_diff);
1.1066 +
1.1067 + } else if (!merge_rename(cur_sim, sux_sim, finished_slot, 0)) {
1.1068 + assert(size_diff > 0, "must be");
1.1069 +
1.1070 + merge_insert_pop(instrs, cur_sim);
1.1071 + size_diff--;
1.1072 + }
1.1073 + assert(cur_sim->stack_size() == 0 || cur_sim->get_slot(0) != reg, "register must have been changed");
1.1074 + }
1.1075 +
1.1076 + while (finished_slot >= 0 && cur_sim->get_slot(finished_slot) == sux_sim->get_slot(finished_slot)) {
1.1077 + finished_slot--;
1.1078 + }
1.1079 +
1.1080 + if (finished_slot >= 0) {
1.1081 + int reg = cur_sim->get_slot(finished_slot);
1.1082 +
1.1083 + if (sux_sim->contains(reg) || !merge_rename(cur_sim, sux_sim, finished_slot, finished_slot)) {
1.1084 + assert(sux_sim->contains(reg) || size_diff > 0, "must be");
1.1085 + merge_insert_xchg(instrs, cur_sim, finished_slot);
1.1086 + }
1.1087 + assert(cur_sim->get_slot(finished_slot) != reg, "register must have been changed");
1.1088 + }
1.1089 + }
1.1090 +
1.1091 +#ifndef PRODUCT
1.1092 + if (TraceFPUStack) {
1.1093 + tty->print("after merging: pred: "); cur_sim->print(); tty->cr();
1.1094 + tty->print(" sux: "); sux_sim->print(); tty->cr();
1.1095 + tty->cr();
1.1096 + }
1.1097 +#endif
1.1098 + assert(cur_sim->stack_size() == sux_sim->stack_size(), "stack size must be equal now");
1.1099 +}
1.1100 +
1.1101 +
1.1102 +void FpuStackAllocator::merge_cleanup_fpu_stack(LIR_List* instrs, FpuStackSim* cur_sim, BitMap& live_fpu_regs) {
1.1103 +#ifndef PRODUCT
1.1104 + if (TraceFPUStack) {
1.1105 + tty->cr();
1.1106 + tty->print("before cleanup: state: "); cur_sim->print(); tty->cr();
1.1107 + tty->print(" live: "); live_fpu_regs.print_on(tty); tty->cr();
1.1108 + }
1.1109 +#endif
1.1110 +
1.1111 + int slot = 0;
1.1112 + while (slot < cur_sim->stack_size()) {
1.1113 + int reg = cur_sim->get_slot(slot);
1.1114 + if (!live_fpu_regs.at(reg)) {
1.1115 + if (slot != 0) {
1.1116 + merge_insert_xchg(instrs, cur_sim, slot);
1.1117 + }
1.1118 + merge_insert_pop(instrs, cur_sim);
1.1119 + } else {
1.1120 + slot++;
1.1121 + }
1.1122 + }
1.1123 +
1.1124 +#ifndef PRODUCT
1.1125 + if (TraceFPUStack) {
1.1126 + tty->print("after cleanup: state: "); cur_sim->print(); tty->cr();
1.1127 + tty->print(" live: "); live_fpu_regs.print_on(tty); tty->cr();
1.1128 + tty->cr();
1.1129 + }
1.1130 +
1.1131 + // check if fpu stack only contains live registers
1.1132 + for (unsigned int i = 0; i < live_fpu_regs.size(); i++) {
1.1133 + if (live_fpu_regs.at(i) != cur_sim->contains(i)) {
1.1134 + tty->print_cr("mismatch between required and actual stack content");
1.1135 + break;
1.1136 + }
1.1137 + }
1.1138 +#endif
1.1139 +}
1.1140 +
1.1141 +
1.1142 +bool FpuStackAllocator::merge_fpu_stack_with_successors(BlockBegin* block) {
1.1143 +#ifndef PRODUCT
1.1144 + if (TraceFPUStack) {
1.1145 + tty->print_cr("Propagating FPU stack state for B%d at LIR_Op position %d to successors:",
1.1146 + block->block_id(), pos());
1.1147 + sim()->print();
1.1148 + tty->cr();
1.1149 + }
1.1150 +#endif
1.1151 +
1.1152 + bool changed = false;
1.1153 + int number_of_sux = block->number_of_sux();
1.1154 +
1.1155 + if (number_of_sux == 1 && block->sux_at(0)->number_of_preds() > 1) {
1.1156 + // The successor has at least two incoming edges, so a stack merge will be necessary
1.1157 + // If this block is the first predecessor, cleanup the current stack and propagate it
1.1158 + // If this block is not the first predecessor, a stack merge will be necessary
1.1159 +
1.1160 + BlockBegin* sux = block->sux_at(0);
1.1161 + intArray* state = sux->fpu_stack_state();
1.1162 + LIR_List* instrs = new LIR_List(_compilation);
1.1163 +
1.1164 + if (state != NULL) {
1.1165 + // Merge with a successors that already has a FPU stack state
1.1166 + // the block must only have one successor because critical edges must been split
1.1167 + FpuStackSim* cur_sim = sim();
1.1168 + FpuStackSim* sux_sim = temp_sim();
1.1169 + sux_sim->read_state(state);
1.1170 +
1.1171 + merge_fpu_stack(instrs, cur_sim, sux_sim);
1.1172 +
1.1173 + } else {
1.1174 + // propagate current FPU stack state to successor without state
1.1175 + // clean up stack first so that there are no dead values on the stack
1.1176 + if (ComputeExactFPURegisterUsage) {
1.1177 + FpuStackSim* cur_sim = sim();
1.1178 + BitMap live_fpu_regs = block->sux_at(0)->fpu_register_usage();
1.1179 + assert(live_fpu_regs.size() == FrameMap::nof_fpu_regs, "missing register usage");
1.1180 +
1.1181 + merge_cleanup_fpu_stack(instrs, cur_sim, live_fpu_regs);
1.1182 + }
1.1183 +
1.1184 + intArray* state = sim()->write_state();
1.1185 + if (TraceFPUStack) {
1.1186 + tty->print_cr("Setting FPU stack state of B%d (merge path)", sux->block_id());
1.1187 + sim()->print(); tty->cr();
1.1188 + }
1.1189 + sux->set_fpu_stack_state(state);
1.1190 + }
1.1191 +
1.1192 + if (instrs->instructions_list()->length() > 0) {
1.1193 + lir()->insert_before(pos(), instrs);
1.1194 + set_pos(instrs->instructions_list()->length() + pos());
1.1195 + changed = true;
1.1196 + }
1.1197 +
1.1198 + } else {
1.1199 + // Propagate unmodified Stack to successors where a stack merge is not necessary
1.1200 + intArray* state = sim()->write_state();
1.1201 + for (int i = 0; i < number_of_sux; i++) {
1.1202 + BlockBegin* sux = block->sux_at(i);
1.1203 +
1.1204 +#ifdef ASSERT
1.1205 + for (int j = 0; j < sux->number_of_preds(); j++) {
1.1206 + assert(block == sux->pred_at(j), "all critical edges must be broken");
1.1207 + }
1.1208 +
1.1209 + // check if new state is same
1.1210 + if (sux->fpu_stack_state() != NULL) {
1.1211 + intArray* sux_state = sux->fpu_stack_state();
1.1212 + assert(state->length() == sux_state->length(), "overwriting existing stack state");
1.1213 + for (int j = 0; j < state->length(); j++) {
1.1214 + assert(state->at(j) == sux_state->at(j), "overwriting existing stack state");
1.1215 + }
1.1216 + }
1.1217 +#endif
1.1218 +#ifndef PRODUCT
1.1219 + if (TraceFPUStack) {
1.1220 + tty->print_cr("Setting FPU stack state of B%d", sux->block_id());
1.1221 + sim()->print(); tty->cr();
1.1222 + }
1.1223 +#endif
1.1224 +
1.1225 + sux->set_fpu_stack_state(state);
1.1226 + }
1.1227 + }
1.1228 +
1.1229 +#ifndef PRODUCT
1.1230 + // assertions that FPU stack state conforms to all successors' states
1.1231 + intArray* cur_state = sim()->write_state();
1.1232 + for (int i = 0; i < number_of_sux; i++) {
1.1233 + BlockBegin* sux = block->sux_at(i);
1.1234 + intArray* sux_state = sux->fpu_stack_state();
1.1235 +
1.1236 + assert(sux_state != NULL, "no fpu state");
1.1237 + assert(cur_state->length() == sux_state->length(), "incorrect length");
1.1238 + for (int i = 0; i < cur_state->length(); i++) {
1.1239 + assert(cur_state->at(i) == sux_state->at(i), "element not equal");
1.1240 + }
1.1241 + }
1.1242 +#endif
1.1243 +
1.1244 + return changed;
1.1245 +}
