truffle: src/share/vm/opto/output.cpp comparison

comparison src/share/vm/opto/output.cpp @ 3851:95134e034042

7063629: use cbcond in C2 generated code on T4 Summary: Use new short branch instruction in C2 generated code. Reviewed-by: never

author	kvn
date	Thu, 11 Aug 2011 12:08:11 -0700
parents	c7b60b601eb4
children	11211f7cb5a0

comparison

equal deleted inserted replaced

-:6987871cfb9b
+:95134e034042
 # ifdef ENABLE_ZAP_DEAD_LOCALS
 if ( ZapDeadCompiledLocals )  Insert_zap_nodes();
 # endif
+uint* blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1);
+blk_starts[0]    = 0;
+// Initialize code buffer and process short branches.
+CodeBuffer* cb = init_buffer(blk_starts);
+if (cb == NULL || failing())  return;
 ScheduleAndBundle();
 #ifndef PRODUCT
 if (trace_opto_output()) {
 tty->print("\n---- After ScheduleAndBundle ----\n");
 }
 #endif
 if (failing())  return;
+finalize_offsets_and_shorten(blk_starts);
 BuildOopMaps();
 if (failing())  return;
-Fill_buffer();
+fill_buffer(cb, blk_starts);
 }
 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
 // Determine if we need to generate a stack overflow check.
 // Do it if the method is not a stub function and
 } // f( b->head()->is_Loop() )
 } // for( i <= last_block )
 } // if( MaxLoopPad < OptoLoopAlignment-1 )
 }
-//----------------------Shorten_branches---------------------------------------
+//----------------------shorten_branches---------------------------------------
 // The architecture description provides short branch variants for some long
 // branch instructions. Replace eligible long branches with short branches.
-void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, int& stub_size) {
+void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
-// fill in the nop array for bundling computations
-MachNode *_nop_list[Bundle::_nop_count];
-Bundle::initialize_nops(_nop_list, this);
 // ------------------
 // Compute size of each block, method size, and relocation information size
-uint *jmp_end    = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks);
+uint nblocks  = _cfg->_num_blocks;
-uint *blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1);
-DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); )
+uint*      jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
-DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); )
+uint*      jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
-blk_starts[0]    = 0;
+int*       jmp_nidx   = NEW_RESOURCE_ARRAY(int ,nblocks);
+DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
+DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
+bool has_short_branch_candidate = false;
 // Initialize the sizes to 0
 code_size  = 0;          // Size in bytes of generated code
 stub_size  = 0;          // Size in bytes of all stub entries
 // Size in bytes of all relocation entries, including those in local stubs.
 // Start with 2-bytes of reloc info for the unvalidated entry point
 reloc_size = 1;          // Number of relocation entries
 // Make three passes.  The first computes pessimistic blk_starts,
-// relative jmp_end and reloc_size information.  The second performs
+// relative jmp_offset and reloc_size information.  The second performs
 // short branch substitution using the pessimistic sizing.  The
 // third inserts nops where needed.
-Node *nj; // tmp
 // Step one, perform a pessimistic sizing pass.
-uint i;
+uint last_call_adr = max_uint;
-uint min_offset_from_last_call = 1;  // init to a positive value
+uint last_avoid_back_to_back_adr = max_uint;
 uint nop_size = (new (this) MachNopNode())->size(_regalloc);
-for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks
+for (uint i = 0; i < nblocks; i++) { // For all blocks
 Block *b = _cfg->_blocks[i];
+// During short branch replacement, we store the relative (to blk_starts)
+// offset of jump in jmp_offset, rather than the absolute offset of jump.
+// This is so that we do not need to recompute sizes of all nodes when
+// we compute correct blk_starts in our next sizing pass.
+jmp_offset[i] = 0;
+jmp_size[i]   = 0;
+jmp_nidx[i]   = -1;
+DEBUG_ONLY( jmp_target[i] = 0; )
+DEBUG_ONLY( jmp_rule[i]   = 0; )
 // Sum all instruction sizes to compute block size
 uint last_inst = b->_nodes.size();
 uint blk_size = 0;
-for( uint j = 0; j<last_inst; j++ ) {
+for (uint j = 0; j < last_inst; j++) {
-nj = b->_nodes[j];
+Node* nj = b->_nodes[j];
 uint inst_size = nj->size(_regalloc);
-blk_size += inst_size;
 // Handle machine instruction nodes
-if( nj->is_Mach() ) {
+if (nj->is_Mach()) {
 MachNode *mach = nj->as_Mach();
 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
 reloc_size += mach->reloc();
 if( mach->is_MachCall() ) {
 MachCallNode *mcall = mach->as_MachCall();
 reloc_size += reloc_java_to_interp();
 }
 } else if (mach->is_MachSafePoint()) {
 // If call/safepoint are adjacent, account for possible
 // nop to disambiguate the two safepoints.
-if (min_offset_from_last_call == 0) {
+// ScheduleAndBundle() can rearrange nodes in a block,
+// check for all offsets inside this block.
+if (last_call_adr >= blk_starts[i]) {
 blk_size += nop_size;
 }
 }
-}
+if (mach->avoid_back_to_back()) {
-min_offset_from_last_call += inst_size;
+// Nop is inserted between "avoid back to back" instructions.
+// ScheduleAndBundle() can rearrange nodes in a block,
+// check for all offsets inside this block.
+if (last_avoid_back_to_back_adr >= blk_starts[i]) {
+blk_size += nop_size;
+}
+}
+if (mach->may_be_short_branch()) {
+if (!nj->is_Branch()) {
+#ifndef PRODUCT
+nj->dump(3);
+#endif
+Unimplemented();
+}
+assert(jmp_nidx[i] == -1, "block should have only one branch");
+jmp_offset[i] = blk_size;
+jmp_size[i]   = inst_size;
+jmp_nidx[i]   = j;
+has_short_branch_candidate = true;
+}
+}
+blk_size += inst_size;
 // Remember end of call offset
 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
-min_offset_from_last_call = 0;
+last_call_adr = blk_starts[i]+blk_size;
 }
-}
+// Remember end of avoid_back_to_back offset
+if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) {
-// During short branch replacement, we store the relative (to blk_starts)
+last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
-// end of jump in jmp_end, rather than the absolute end of jump.  This
+}
-// is so that we do not need to recompute sizes of all nodes when we compute
+}
-// correct blk_starts in our next sizing pass.
-jmp_end[i] = blk_size;
-DEBUG_ONLY( jmp_target[i] = 0; )
 // When the next block starts a loop, we may insert pad NOP
 // instructions.  Since we cannot know our future alignment,
 // assume the worst.
-if( i<_cfg->_num_blocks-1 ) {
+if (i< nblocks-1) {
 Block *nb = _cfg->_blocks[i+1];
 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
-if( max_loop_pad > 0 ) {
+if (max_loop_pad > 0) {
 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
 blk_size += max_loop_pad;
 }
 }
 // Save block size; update total method size
 blk_starts[i+1] = blk_starts[i]+blk_size;
 }
 // Step two, replace eligible long jumps.
+bool progress = true;
-// Note: this will only get the long branches within short branch
+uint last_may_be_short_branch_adr = max_uint;
-//   range. Another pass might detect more branches that became
+while (has_short_branch_candidate && progress) {
-//   candidates because the shortening in the first pass exposed
+progress = false;
-//   more opportunities. Unfortunately, this would require
+has_short_branch_candidate = false;
-//   recomputing the starting and ending positions for the blocks
+int adjust_block_start = 0;
-for( i=0; i<_cfg->_num_blocks; i++ ) {
+for (uint i = 0; i < nblocks; i++) {
 Block *b = _cfg->_blocks[i];
+int idx = jmp_nidx[i];
-int j;
+MachNode* mach = (idx == -1) ? NULL: b->_nodes[idx]->as_Mach();
-// Find the branch; ignore trailing NOPs.
+if (mach != NULL && mach->may_be_short_branch()) {
-for( j = b->_nodes.size()-1; j>=0; j-- ) {
+#ifdef ASSERT
-nj = b->_nodes[j];
+assert(jmp_size[i] > 0 && mach->is_Branch(), "sanity");
-if( !nj->is_Mach() || nj->as_Mach()->ideal_Opcode() != Op_Con )
+int j;
-break;
+// Find the branch; ignore trailing NOPs.
-}
+for (j = b->_nodes.size()-1; j>=0; j--) {
+Node* n = b->_nodes[j];
-if (j >= 0) {
+if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
-if( nj->is_Mach() && nj->as_Mach()->may_be_short_branch() ) {
+break;
-MachNode *mach = nj->as_Mach();
+}
+assert(j >= 0 && j == idx && b->_nodes[j] == (Node*)mach, "sanity");
+#endif
+int br_size = jmp_size[i];
+int br_offs = blk_starts[i] + jmp_offset[i];
 // This requires the TRUE branch target be in succs[0]
 uint bnum = b->non_connector_successor(0)->_pre_order;
-uintptr_t target = blk_starts[bnum];
+int offset = blk_starts[bnum] - br_offs;
-if( mach->is_Branch() ) {
+if (bnum > i) { // adjust following block's offset
-int offset = target-(blk_starts[i] + jmp_end[i]);
+offset -= adjust_block_start;
-if (_matcher->is_short_branch_offset(mach->rule(), offset)) {
+}
-// We've got a winner.  Replace this branch.
+// In the following code a nop could be inserted before
-MachNode* replacement = mach->short_branch_version(this);
+// the branch which will increase the backward distance.
-b->_nodes.map(j, replacement);
+bool needs_padding = ((uint)br_offs == last_may_be_short_branch_adr);
-mach->subsume_by(replacement);
+if (needs_padding && offset <= 0)
+offset -= nop_size;
-// Update the jmp_end size to save time in our
-// next pass.
+if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
-jmp_end[i] -= (mach->size(_regalloc) - replacement->size(_regalloc));
+// We've got a winner.  Replace this branch.
-DEBUG_ONLY( jmp_target[i] = bnum; );
+MachNode* replacement = mach->short_branch_version(this);
-DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
+// Update the jmp_size.
+int new_size = replacement->size(_regalloc);
+int diff     = br_size - new_size;
+assert(diff >= (int)nop_size, "short_branch size should be smaller");
+// Conservatively take into accound padding between
+// avoid_back_to_back branches. Previous branch could be
+// converted into avoid_back_to_back branch during next
+// rounds.
+if (needs_padding && replacement->avoid_back_to_back()) {
+jmp_offset[i] += nop_size;
+diff -= nop_size;
 }
+adjust_block_start += diff;
+b->_nodes.map(idx, replacement);
+mach->subsume_by(replacement);
+mach = replacement;
+progress = true;
+jmp_size[i] = new_size;
+DEBUG_ONLY( jmp_target[i] = bnum; );
+DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
 } else {
-#ifndef PRODUCT
+// The jump distance is not short, try again during next iteration.
-mach->dump(3);
+has_short_branch_candidate = true;
-#endif
-Unimplemented();
 }
-}
+} // (mach->may_be_short_branch())
-}
+if (mach != NULL && (mach->may_be_short_branch() ||
-}
+mach->avoid_back_to_back())) {
+last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
+}
+blk_starts[i+1] -= adjust_block_start;
+}
+}
+#ifdef ASSERT
+for (uint i = 0; i < nblocks; i++) { // For all blocks
+if (jmp_target[i] != 0) {
+int br_size = jmp_size[i];
+int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
+if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
+tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
+}
+assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
+}
+}
+#endif
+// Step 3, compute the offsets of all blocks, will be done in finalize_offsets_and_shorten()
+// after ScheduleAndBundle().
+// ------------------
+// Compute size for code buffer
+code_size = blk_starts[nblocks];
+// Relocation records
+reloc_size += 1;              // Relo entry for exception handler
+// Adjust reloc_size to number of record of relocation info
+// Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
+// a relocation index.
+// The CodeBuffer will expand the locs array if this estimate is too low.
+reloc_size *= 10 / sizeof(relocInfo);
+}
+//----------------------finalize_offsets_and_shorten-------------------------
+void Compile::finalize_offsets_and_shorten(uint* blk_starts) {
+// blk_starts[] contains offsets calculated during short branches processing,
+// offsets should not be increased during following steps.
 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
 // of a loop. It is used to determine the padding for loop alignment.
 compute_loop_first_inst_sizes();
-// Step 3, compute the offsets of all the labels
+uint nblocks  = _cfg->_num_blocks;
+#ifdef ASSERT
+uint*      jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
+uint*      jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
+uint*      jmp_size   = NEW_RESOURCE_ARRAY(uint,nblocks);
+uint*      jmp_rule   = NEW_RESOURCE_ARRAY(uint,nblocks);
+#endif
+// Inserts nops where needed and do final short branches replacement.
+uint nop_size = (new (this) MachNopNode())->size(_regalloc);
 uint last_call_adr = max_uint;
-for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks
+uint last_avoid_back_to_back_adr = max_uint;
-// copy the offset of the beginning to the corresponding label
-assert(labels[i].is_unused(), "cannot patch at this point");
+assert(blk_starts[0] == 0, "sanity");
-labels[i].bind_loc(blk_starts[i], CodeBuffer::SECT_INSTS);
+uint current_offset = 0;
+uint block_alignment_padding = 0;
-// insert padding for any instructions that need it
+for (uint i=0; i < nblocks; i++) { // For all blocks
 Block *b = _cfg->_blocks[i];
+#ifdef ASSERT
+jmp_target[i] = 0;
+jmp_offset[i] = 0;
+jmp_size[i]   = 0;
+jmp_rule[i]   = 0;
+#endif
+// Maximum alignment was added before loop block during
+// Step One, as result padding for nodes was not added.
+// Take this into account for block's size change check
+// and allow increase block's size by the difference
+// of maximum and actual alignment paddings.
+DEBUG_ONLY( uint orig_blk_size = blk_starts[i+1] - blk_starts[i] + block_alignment_padding; )
+uint blk_offset = current_offset;
 uint last_inst = b->_nodes.size();
-uint adr = blk_starts[i];
+for (uint j = 0; j<last_inst; j++) {
-for( uint j = 0; j<last_inst; j++ ) {
+Node* nj = b->_nodes[j];
-nj = b->_nodes[j];
-if( nj->is_Mach() ) {
+if (valid_bundle_info(nj) &&
-int padding = nj->as_Mach()->compute_padding(adr);
+node_bundling(nj)->used_in_unconditional_delay()) {
+continue; // Skip instruction in delay slot
+}
+uint inst_size = nj->size(_regalloc);
+if (nj->is_Mach()) {
+MachNode *mach = nj->as_Mach();
+int padding = mach->compute_padding(current_offset);
 // If call/safepoint are adjacent insert a nop (5010568)
 if (padding == 0 && nj->is_MachSafePoint() && !nj->is_MachCall() &&
-adr == last_call_adr ) {
+current_offset == last_call_adr) {
 padding = nop_size;
 }
-if(padding > 0) {
+// Inserted a nop between "avoid back to back" instructions.
+if (padding == 0 && mach->avoid_back_to_back() &&
+current_offset == last_avoid_back_to_back_adr) {
+padding = nop_size;
+}
+if (padding > 0) {
 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
 int nops_cnt = padding / nop_size;
 MachNode *nop = new (this) MachNopNode(nops_cnt);
 b->_nodes.insert(j++, nop);
-_cfg->_bbs.map( nop->_idx, b );
+_cfg->_bbs.map(nop->_idx, b);
-adr += padding;
 last_inst++;
+current_offset += padding;
 }
-}
-adr += nj->size(_regalloc);
+// Try to replace long branch if delay slot is not used,
+// it is mostly for back branches since forward branch's
+// distance is not updated yet.
+bool delay_slot_is_used = valid_bundle_info(nj) &&
+node_bundling(nj)->use_unconditional_delay();
+if (!delay_slot_is_used && mach->may_be_short_branch()) {
+int br_size = inst_size;
+// This requires the TRUE branch target be in succs[0]
+uint bnum = b->non_connector_successor(0)->_pre_order;
+int offset = blk_starts[bnum] - current_offset;
+if (bnum >= i) {
+// Current and following block's offset are not
+// finilized yet, adjust distance.
+offset -= (blk_starts[i] - blk_offset);
+}
+// In the following code a nop could be inserted before
+// the branch which will increase the backward distance.
+bool needs_padding = (current_offset == last_avoid_back_to_back_adr);
+if (needs_padding && offset <= 0)
+offset -= nop_size;
+if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
+// We've got a winner.  Replace this branch.
+MachNode* replacement = mach->short_branch_version(this);
+// Update the jmp_size.
+int new_size = replacement->size(_regalloc);
+assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
+// Conservatively take into accound padding between
+// avoid_back_to_back branches. Previous branch could be
+// converted into avoid_back_to_back branch during next
+// rounds.
+if (needs_padding && replacement->avoid_back_to_back()) {
+MachNode *nop = new (this) MachNopNode();
+b->_nodes.insert(j++, nop);
+_cfg->_bbs.map(nop->_idx, b);
+last_inst++;
+current_offset += nop_size;
+}
+inst_size = new_size;
+b->_nodes.map(j, replacement);
+mach->subsume_by(replacement);
+nj = replacement;
+#ifdef ASSERT
+jmp_target[i] = bnum;
+jmp_offset[i] = current_offset - blk_offset;
+jmp_size[i]   = new_size;
+jmp_rule[i]   = mach->rule();
+#endif
+}
+}
+}
+current_offset += inst_size;
 // Remember end of call offset
 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
-last_call_adr = adr;
+last_call_adr = current_offset;
 }
-}
+// Remember end of avoid_back_to_back offset
+if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) {
-if ( i != _cfg->_num_blocks-1) {
+last_avoid_back_to_back_adr = current_offset;
-// Get the size of the block
+}
-uint blk_size = adr - blk_starts[i];
+}
+assert(blk_offset <= blk_starts[i], "shouldn't increase distance");
-// When the next block is the top of a loop, we may insert pad NOP
+blk_starts[i] = blk_offset;
-// instructions.
+// When the next block is the top of a loop, we may insert pad NOP
+// instructions.
+if (i < nblocks-1) {
 Block *nb = _cfg->_blocks[i+1];
-int current_offset = blk_starts[i] + blk_size;
+int padding = nb->alignment_padding(current_offset);
-current_offset += nb->alignment_padding(current_offset);
+if (padding > 0) {
-// Save block size; update total method size
+assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
-blk_starts[i+1] = current_offset;
+int nops_cnt = padding / nop_size;
-}
+MachNode *nop = new (this) MachNopNode(nops_cnt);
-}
+b->_nodes.insert(b->_nodes.size(), nop);
+_cfg->_bbs.map(nop->_idx, b);
+current_offset += padding;
+}
+int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
+assert(max_loop_pad >= padding, "sanity");
+block_alignment_padding = max_loop_pad - padding;
+}
+assert(orig_blk_size >= (current_offset - blk_offset), "shouldn't increase block size");
+}
+blk_starts[nblocks] = current_offset;
 #ifdef ASSERT
-for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks
+for (uint i = 0; i < nblocks; i++) { // For all blocks
-if( jmp_target[i] != 0 ) {
+if (jmp_target[i] != 0) {
-int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_end[i]);
+int br_size = jmp_size[i];
-if (!_matcher->is_short_branch_offset(jmp_rule[i], offset)) {
+int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
-tty->print_cr("target (%d) - jmp_end(%d) = offset (%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_end[i], offset, i, jmp_target[i]);
+if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
-}
+tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
-assert(_matcher->is_short_branch_offset(jmp_rule[i], offset), "Displacement too large for short jmp");
+}
-}
+assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
 }
-#endif
+}
+#endif
-// ------------------
-// Compute size for code buffer
-code_size   = blk_starts[i-1] + jmp_end[i-1];
-// Relocation records
-reloc_size += 1;              // Relo entry for exception handler
-// Adjust reloc_size to number of record of relocation info
-// Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
-// a relocation index.
-// The CodeBuffer will expand the locs array if this estimate is too low.
-reloc_size   *= 10 / sizeof(relocInfo);
 }
 //------------------------------FillLocArray-----------------------------------
 // Create a bit of debug info and append it to the array.  The mapping is from
 // Java local or expression stack to constant, register or stack-slot.  For
 debug_info->end_non_safepoint(pc_offset);
 }
-// helper for Fill_buffer bailout logic
+// helper for fill_buffer bailout logic
 static void turn_off_compiler(Compile* C) {
 if (CodeCache::largest_free_block() >= CodeCacheMinimumFreeSpace*10) {
 // Do not turn off compilation if a single giant method has
 // blown the code cache size.
 C->record_failure("excessive request to CodeCache");
 C->record_failure("CodeCache is full");
 }
 }
-//------------------------------Fill_buffer------------------------------------
+//------------------------------init_buffer------------------------------------
-void Compile::Fill_buffer() {
+CodeBuffer* Compile::init_buffer(uint* blk_starts) {
 // Set the initially allocated size
 int  code_req   = initial_code_capacity;
 int  locs_req   = initial_locs_capacity;
 int  stub_req   = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
 int  const_req  = initial_const_capacity;
-bool labels_not_set = true;
 int  pad_req    = NativeCall::instruction_size;
 // The extra spacing after the code is necessary on some platforms.
 // Sometimes we need to patch in a jump after the last instruction,
 // if the nmethod has been deoptimized.  (See 4932387, 4894843.)
-uint i;
 // Compute the byte offset where we can store the deopt pc.
 if (fixed_slots() != 0) {
 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
 }
 // get rid of this hack and have SpillCopy generate stfspill/ldffill
 // instead of stfd/stfs/ldfd/ldfs.
 _frame_slots += 8*(16/BytesPerInt);
 }
 #endif
-assert( _frame_slots >= 0 && _frame_slots < 1000000, "sanity check" );
+assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
-// Create an array of unused labels, one for each basic block
-Label *blk_labels = NEW_RESOURCE_ARRAY(Label, _cfg->_num_blocks+1);
-for( i=0; i <= _cfg->_num_blocks; i++ ) {
-blk_labels[i].init();
-}
 if (has_mach_constant_base_node()) {
 // Fill the constant table.
-// Note:  This must happen before Shorten_branches.
+// Note:  This must happen before shorten_branches.
-for (i = 0; i < _cfg->_num_blocks; i++) {
+for (uint i = 0; i < _cfg->_num_blocks; i++) {
 Block* b = _cfg->_blocks[i];
 for (uint j = 0; j < b->_nodes.size(); j++) {
 Node* n = b->_nodes[j];
 }
 // Initialize the space for the BufferBlob used to find and verify
 // instruction size in MachNode::emit_size()
 init_scratch_buffer_blob(const_req);
-if (failing())  return; // Out of memory
+if (failing())  return NULL; // Out of memory
-// If this machine supports different size branch offsets, then pre-compute
+// Pre-compute the length of blocks and replace
-// the length of the blocks
+// long branches with short if machine supports it.
-if( _matcher->is_short_branch_offset(-1, 0) ) {
+shorten_branches(blk_starts, code_req, locs_req, stub_req);
-Shorten_branches(blk_labels, code_req, locs_req, stub_req);
-labels_not_set = false;
-}
 // nmethod and CodeBuffer count stubs & constants as part of method's code.
 int exception_handler_req = size_exception_handler();
 int deopt_handler_req = size_deopt_handler();
 exception_handler_req += MAX_stubs_size; // add marginal slop for handler
 cb->initialize(total_req, locs_req);
 // Have we run out of code space?
 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
 turn_off_compiler(this);
-return;
+return NULL;
 }
 // Configure the code buffer.
 cb->initialize_consts_size(const_req);
 cb->initialize_stubs_size(stub_req);
 cb->initialize_oop_recorder(env()->oop_recorder());
 // fill in the nop array for bundling computations
 MachNode *_nop_list[Bundle::_nop_count];
 Bundle::initialize_nops(_nop_list, this);
+return cb;
+}
+//------------------------------fill_buffer------------------------------------
+void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
 // Create oopmap set.
 _oop_map_set = new OopMapSet();
 // !!!!! This preserves old handling of oopmaps for now
 debug_info()->set_oopmaps(_oop_map_set);
 uint  return_offset = 0;
 int nop_size = (new (this) MachNopNode())->size(_regalloc);
 int previous_offset = 0;
 int current_offset  = 0;
+#ifdef ASSERT
 int last_call_offset = -1;
+int last_avoid_back_to_back_offset = -1;
+#endif
 // Create an array of unused labels, one for each basic block, if printing is enabled
 #ifndef PRODUCT
 int *node_offsets      = NULL;
-uint  node_offset_limit = unique();
+uint node_offset_limit = unique();
+if (print_assembly())
-if ( print_assembly() )
 node_offsets         = NEW_RESOURCE_ARRAY(int, node_offset_limit);
 #endif
 NonSafepointEmitter non_safepoints(this);  // emit non-safepoints lazily
 // Emit the constant table.
 if (has_mach_constant_base_node()) {
 constant_table().emit(*cb);
 }
+// Create an array of labels, one for each basic block
+Label *blk_labels = NEW_RESOURCE_ARRAY(Label, _cfg->_num_blocks+1);
+for (uint i=0; i <= _cfg->_num_blocks; i++) {
+blk_labels[i].init();
+}
 // ------------------
 // Now fill in the code buffer
 Node *delay_slot = NULL;
-for( i=0; i < _cfg->_num_blocks; i++ ) {
+for (uint i=0; i < _cfg->_num_blocks; i++) {
+guarantee(blk_starts[i] == (uint)cb->insts_size(),"should not change size");
 Block *b = _cfg->_blocks[i];
 Node *head = b->head();
 // If this block needs to start aligned (i.e, can be reached other
 // than by falling-thru from the previous block), then force the
 // start of a new bundle.
-if( Pipeline::requires_bundling() && starts_bundle(head) )
+if (Pipeline::requires_bundling() && starts_bundle(head))
 cb->flush_bundle(true);
+#ifdef ASSERT
+if (!b->is_connector()) {
+stringStream st;
+b->dump_head(&_cfg->_bbs, &st);
+MacroAssembler(cb).block_comment(st.as_string());
+}
+#endif
 // Define the label at the beginning of the basic block
-if (labels_not_set) {
+MacroAssembler(cb).bind(blk_labels[b->_pre_order]);
-MacroAssembler(cb).bind(blk_labels[b->_pre_order]);
-} else {
-assert(blk_labels[b->_pre_order].loc_pos() == cb->insts_size(),
-err_msg("label position does not match code offset: %d != %d",
-blk_labels[b->_pre_order].loc_pos(), cb->insts_size()));
-}
 uint last_inst = b->_nodes.size();
 // Emit block normally, except for last instruction.
 // Emit means "dump code bits into code buffer".
-for( uint j = 0; j<last_inst; j++ ) {
+for (uint j = 0; j<last_inst; j++) {
 // Get the node
 Node* n = b->_nodes[j];
 // See if delay slots are supported
 continue;
 }
 // If this starts a new instruction group, then flush the current one
 // (but allow split bundles)
-if( Pipeline::requires_bundling() && starts_bundle(n) )
+if (Pipeline::requires_bundling() && starts_bundle(n))
 cb->flush_bundle(false);
 // The following logic is duplicated in the code ifdeffed for
 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file.  It
 // should be factored out.  Or maybe dispersed to the nodes?
 // Special handling for SafePoint/Call Nodes
 bool is_mcall = false;
-if( n->is_Mach() ) {
+if (n->is_Mach()) {
 MachNode *mach = n->as_Mach();
 is_mcall = n->is_MachCall();
 bool is_sfn = n->is_MachSafePoint();
 // If this requires all previous instructions be flushed, then do so
-if( is_sfn || is_mcall || mach->alignment_required() != 1) {
+if (is_sfn || is_mcall || mach->alignment_required() != 1) {
 cb->flush_bundle(true);
 current_offset = cb->insts_size();
 }
+#ifdef ASSERT
+// A padding may be needed again since a previous instruction
+// could be moved to delay slot.
 // align the instruction if necessary
 int padding = mach->compute_padding(current_offset);
 // Make sure safepoint node for polling is distinct from a call's
 // return by adding a nop if needed.
-if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset ) {
+if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
 padding = nop_size;
 }
-assert( labels_not_set || padding == 0, "instruction should already be aligned");
+if (padding == 0 && mach->avoid_back_to_back() &&
+current_offset == last_avoid_back_to_back_offset) {
-if(padding > 0) {
+// Avoid back to back some instructions.
-assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
+padding = nop_size;
-int nops_cnt = padding / nop_size;
-MachNode *nop = new (this) MachNopNode(nops_cnt);
-b->_nodes.insert(j++, nop);
-last_inst++;
-_cfg->_bbs.map( nop->_idx, b );
-nop->emit(*cb, _regalloc);
-cb->flush_bundle(true);
-current_offset = cb->insts_size();
 }
+assert(padding == 0, "padding should be added already");
+#endif
 // Remember the start of the last call in a basic block
 if (is_mcall) {
 MachCallNode *mcall = mach->as_MachCall();
 // This destination address is NOT PC-relative
 is_sfn = false;
 }
 }
 // sfn will be valid whenever mcall is valid now because of inheritance
-if( is_sfn || is_mcall ) {
+if (is_sfn || is_mcall) {
 // Handle special safepoint nodes for synchronization
-if( !is_mcall ) {
+if (!is_mcall) {
 MachSafePointNode *sfn = mach->as_MachSafePoint();
 // !!!!! Stubs only need an oopmap right now, so bail out
-if( sfn->jvms()->method() == NULL) {
+if (sfn->jvms()->method() == NULL) {
 // Write the oopmap directly to the code blob??!!
 #             ifdef ENABLE_ZAP_DEAD_LOCALS
 assert( !is_node_getting_a_safepoint(sfn),  "logic does not match; false positive");
 #             endif
 continue;
 else if( mach->is_MachNullCheck() ) {
 inct_starts[inct_cnt++] = previous_offset;
 }
 // If this is a branch, then fill in the label with the target BB's label
-else if ( mach->is_Branch() ) {
+else if (mach->is_Branch()) {
-if ( mach->ideal_Opcode() == Op_Jump ) {
+if (mach->ideal_Opcode() == Op_Jump) {
-for (uint h = 0; h < b->_num_succs; h++ ) {
+for (uint h = 0; h < b->_num_succs; h++) {
 Block* succs_block = b->_succs[h];
 for (uint j = 1; j < succs_block->num_preds(); j++) {
 Node* jpn = succs_block->pred(j);
-if ( jpn->is_JumpProj() && jpn->in(0) == mach ) {
+if (jpn->is_JumpProj() && jpn->in(0) == mach) {
 uint block_num = succs_block->non_connector()->_pre_order;
 Label *blkLabel = &blk_labels[block_num];
 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
 }
 }
 }
 }
 #ifdef ASSERT
 // Check that oop-store precedes the card-mark
-else if( mach->ideal_Opcode() == Op_StoreCM ) {
+else if (mach->ideal_Opcode() == Op_StoreCM) {
 uint storeCM_idx = j;
 int count = 0;
 for (uint prec = mach->req(); prec < mach->len(); prec++) {
 Node *oop_store = mach->in(prec);  // Precedence edge
 if (oop_store == NULL) continue;
 }
 assert(count > 0, "storeCM expects at least one precedence edge");
 }
 #endif
-else if( !n->is_Proj() ) {
+else if (!n->is_Proj()) {
 // Remember the beginning of the previous instruction, in case
 // it's followed by a flag-kill and a null-check.  Happens on
 // Intel all the time, with add-to-memory kind of opcodes.
 previous_offset = current_offset;
 }
 return;
 }
 // Save the offset for the listing
 #ifndef PRODUCT
-if( node_offsets && n->_idx < node_offset_limit )
+if (node_offsets && n->_idx < node_offset_limit)
 node_offsets[n->_idx] = cb->insts_size();
 #endif
 // "Normal" instruction case
+DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
 n->emit(*cb, _regalloc);
 current_offset  = cb->insts_size();
+#ifdef ASSERT
+if (n->size(_regalloc) != (current_offset-instr_offset)) {
+n->dump();
+assert(n->size(_regalloc) == (current_offset-instr_offset), "wrong size of mach node");
+}
+#endif
 non_safepoints.observe_instruction(n, current_offset);
+#ifdef ASSERT
 // mcall is last "call" that can be a safepoint
 // record it so we can see if a poll will directly follow it
 // in which case we'll need a pad to make the PcDesc sites unique
 // see  5010568. This can be slightly inaccurate but conservative
 // in the case that return address is not actually at current_offset.
 if (is_mcall) {
 last_call_offset = current_offset;
 }
+if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) {
+// Avoid back to back some instructions.
+last_avoid_back_to_back_offset = current_offset;
+}
+#endif
 // See if this instruction has a delay slot
-if ( valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
+if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
 assert(delay_slot != NULL, "expecting delay slot node");
 // Back up 1 instruction
 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
 // Save the offset for the listing
 #ifndef PRODUCT
-if( node_offsets && delay_slot->_idx < node_offset_limit )
+if (node_offsets && delay_slot->_idx < node_offset_limit)
 node_offsets[delay_slot->_idx] = cb->insts_size();
 #endif
 // Support a SafePoint in the delay slot
-if( delay_slot->is_MachSafePoint() ) {
+if (delay_slot->is_MachSafePoint()) {
 MachNode *mach = delay_slot->as_Mach();
 // !!!!! Stubs only need an oopmap right now, so bail out
-if( !mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL ) {
+if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
 // Write the oopmap directly to the code blob??!!
 #           ifdef ENABLE_ZAP_DEAD_LOCALS
 assert( !is_node_getting_a_safepoint(mach),  "logic does not match; false positive");
 #           endif
 delay_slot = NULL;
 // Don't reuse it
 delay_slot = NULL;
 }
 } // End for all instructions in block
+#ifdef ASSERT
 // If the next block is the top of a loop, pad this block out to align
 // the loop top a little. Helps prevent pipe stalls at loop back branches.
-if( i<_cfg->_num_blocks-1 ) {
+if (i < _cfg->_num_blocks-1) {
 Block *nb = _cfg->_blocks[i+1];
 uint padding = nb->alignment_padding(current_offset);
-if( padding > 0 ) {
+assert(padding == 0, "alignment should be added already");
-MachNode *nop = new (this) MachNopNode(padding / nop_size);
+}
-b->_nodes.insert( b->_nodes.size(), nop );
+#endif
-_cfg->_bbs.map( nop->_idx, b );
-nop->emit(*cb, _regalloc);
-current_offset = cb->insts_size();
-}
-}
 } // End of for all blocks
 non_safepoints.flush_at_end();
 // Offset too large?
 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
 // Create a data structure for all the scheduling information
 Scheduling scheduling(Thread::current()->resource_area(), *this);
-// Initialize the space for the BufferBlob used to find and verify
-// instruction size in MachNode::emit_size()
-init_scratch_buffer_blob(MAX_const_size);
-if (failing())  return;  // Out of memory
 // Walk backwards over each basic block, computing the needed alignment
 // Walk over all the basic blocks
 scheduling.DoScheduling();
 }
 // normally schedule conditional branches (despite them being forced last
 // in the block), because they have delay slots we can fill.  Calls all
 // have their delay slots filled in the template expansions, so we don't
 // bother scheduling them.
 Node *last = bb->_nodes[_bb_end];
+// Ignore trailing NOPs.
+while (_bb_end > 0 && last->is_Mach() &&
+last->as_Mach()->ideal_Opcode() == Op_Con) {
+last = bb->_nodes[--_bb_end];
+}
+assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
 if( last->is_Catch() ||
 // Exclude unreachable path case when Halt node is in a separate block.
 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
 // There must be a prior call.  Skip it.
 while( !bb->_nodes[--_bb_end]->is_MachCall() ) {
 }
 } else {
 // Get DEF'd registers the normal way
 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
+}
+// Kill projections on a branch should appear to occur on the
+// branch, not afterwards, so grab the masks from the projections
+// and process them.
+if (n->is_Branch()) {
+for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
+Node* use = n->fast_out(i);
+if (use->is_Proj()) {
+RegMask rm = use->out_RegMask();// Make local copy
+while( rm.is_NotEmpty() ) {
+OptoReg::Name kill = rm.find_first_elem();
+rm.Remove(kill);
+anti_do_def( b, n, kill, false );
+}
+}
+}
 }
 // Check each register used by this instruction for a following DEF/KILL
 // that must occur afterward and requires an anti-dependence edge.
 for( uint j=0; j<n->req(); j++ ) {

Mercurial > hg > truffle

comparison src/share/vm/opto/output.cpp @ 3851:95134e034042