Mercurial > hg > truffle
annotate src/share/vm/opto/reg_split.cpp @ 12023:d1034bd8cefc
8022284: Hide internal data structure in PhaseCFG
Summary: Hide private node to block mapping using public interface
Reviewed-by: kvn, roland
| author | adlertz |
|---|---|
| date | Wed, 07 Aug 2013 17:56:19 +0200 |
| parents | b274ac1dbe11 |
| children | adb9a7d94cb5 |
| rev | line source |
|---|---|
| 0 | 1 /* |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
2 * Copyright (c) 2000, 2012, Oracle and/or its affiliates. All rights reserved. |
| 0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
| 4 * | |
| 5 * This code is free software; you can redistribute it and/or modify it | |
| 6 * under the terms of the GNU General Public License version 2 only, as | |
| 7 * published by the Free Software Foundation. | |
| 8 * | |
| 9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 12 * version 2 for more details (a copy is included in the LICENSE file that | |
| 13 * accompanied this code). | |
| 14 * | |
| 15 * You should have received a copy of the GNU General Public License version | |
| 16 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 18 * | |
|
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
628
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
|
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
628
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
|
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
628
diff
changeset
|
21 * questions. |
| 0 | 22 * |
| 23 */ | |
| 24 | |
| 1972 | 25 #include "precompiled.hpp" |
| 26 #include "libadt/vectset.hpp" | |
| 27 #include "memory/allocation.inline.hpp" | |
| 28 #include "opto/addnode.hpp" | |
| 29 #include "opto/c2compiler.hpp" | |
| 30 #include "opto/callnode.hpp" | |
| 31 #include "opto/cfgnode.hpp" | |
| 32 #include "opto/chaitin.hpp" | |
| 33 #include "opto/loopnode.hpp" | |
| 34 #include "opto/machnode.hpp" | |
| 0 | 35 |
| 36 //------------------------------Split-------------------------------------- | |
| 605 | 37 // Walk the graph in RPO and for each lrg which spills, propagate reaching |
| 38 // definitions. During propagation, split the live range around regions of | |
| 0 | 39 // High Register Pressure (HRP). If a Def is in a region of Low Register |
| 40 // Pressure (LRP), it will not get spilled until we encounter a region of | |
| 41 // HRP between it and one of its uses. We will spill at the transition | |
| 42 // point between LRP and HRP. Uses in the HRP region will use the spilled | |
| 43 // Def. The first Use outside the HRP region will generate a SpillCopy to | |
| 44 // hoist the live range back up into a register, and all subsequent uses | |
| 45 // will use that new Def until another HRP region is encountered. Defs in | |
| 46 // HRP regions will get trailing SpillCopies to push the LRG down into the | |
| 47 // stack immediately. | |
| 48 // | |
| 49 // As a side effect, unlink from (hence make dead) coalesced copies. | |
| 50 // | |
| 51 | |
| 52 static const char out_of_nodes[] = "out of nodes during split"; | |
| 53 | |
|
10395
b274ac1dbe11
8005956: C2: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG defined in this block
adlertz
parents:
10111
diff
changeset
|
54 static bool contains_no_live_range_input(const Node* def) { |
|
b274ac1dbe11
8005956: C2: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG defined in this block
adlertz
parents:
10111
diff
changeset
|
55 for (uint i = 1; i < def->req(); ++i) { |
|
b274ac1dbe11
8005956: C2: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG defined in this block
adlertz
parents:
10111
diff
changeset
|
56 if (def->in(i) != NULL && def->in_RegMask(i).is_NotEmpty()) { |
|
b274ac1dbe11
8005956: C2: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG defined in this block
adlertz
parents:
10111
diff
changeset
|
57 return false; |
|
b274ac1dbe11
8005956: C2: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG defined in this block
adlertz
parents:
10111
diff
changeset
|
58 } |
|
b274ac1dbe11
8005956: C2: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG defined in this block
adlertz
parents:
10111
diff
changeset
|
59 } |
|
b274ac1dbe11
8005956: C2: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG defined in this block
adlertz
parents:
10111
diff
changeset
|
60 return true; |
|
b274ac1dbe11
8005956: C2: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG defined in this block
adlertz
parents:
10111
diff
changeset
|
61 } |
|
b274ac1dbe11
8005956: C2: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG defined in this block
adlertz
parents:
10111
diff
changeset
|
62 |
| 0 | 63 //------------------------------get_spillcopy_wide----------------------------- |
| 64 // Get a SpillCopy node with wide-enough masks. Use the 'wide-mask', the | |
| 65 // wide ideal-register spill-mask if possible. If the 'wide-mask' does | |
| 66 // not cover the input (or output), use the input (or output) mask instead. | |
| 67 Node *PhaseChaitin::get_spillcopy_wide( Node *def, Node *use, uint uidx ) { | |
| 68 // If ideal reg doesn't exist we've got a bad schedule happening | |
| 69 // that is forcing us to spill something that isn't spillable. | |
| 70 // Bail rather than abort | |
| 71 int ireg = def->ideal_reg(); | |
| 72 if( ireg == 0 || ireg == Op_RegFlags ) { | |
|
415
4d9884b01ba6
6754519: don't emit flag fixup for NaN when condition being tested doesn't need it
never
parents:
337
diff
changeset
|
73 assert(false, "attempted to spill a non-spillable item"); |
| 0 | 74 C->record_method_not_compilable("attempted to spill a non-spillable item"); |
| 75 return NULL; | |
| 76 } | |
| 77 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { | |
| 78 return NULL; | |
| 79 } | |
| 80 const RegMask *i_mask = &def->out_RegMask(); | |
| 81 const RegMask *w_mask = C->matcher()->idealreg2spillmask[ireg]; | |
| 82 const RegMask *o_mask = use ? &use->in_RegMask(uidx) : w_mask; | |
| 83 const RegMask *w_i_mask = w_mask->overlap( *i_mask ) ? w_mask : i_mask; | |
| 84 const RegMask *w_o_mask; | |
| 85 | |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
86 int num_regs = RegMask::num_registers(ireg); |
|
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
87 bool is_vect = RegMask::is_vector(ireg); |
| 0 | 88 if( w_mask->overlap( *o_mask ) && // Overlap AND |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
89 ((num_regs == 1) // Single use or aligned |
|
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
90 || is_vect // or vector |
|
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
91 || !is_vect && o_mask->is_aligned_pairs()) ) { |
|
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
92 assert(!is_vect || o_mask->is_aligned_sets(num_regs), "vectors are aligned"); |
| 0 | 93 // Don't come here for mis-aligned doubles |
| 94 w_o_mask = w_mask; | |
| 95 } else { // wide ideal mask does not overlap with o_mask | |
| 96 // Mis-aligned doubles come here and XMM->FPR moves on x86. | |
| 97 w_o_mask = o_mask; // Must target desired registers | |
| 98 // Does the ideal-reg-mask overlap with o_mask? I.e., can I use | |
| 99 // a reg-reg move or do I need a trip across register classes | |
| 100 // (and thus through memory)? | |
| 101 if( !C->matcher()->idealreg2regmask[ireg]->overlap( *o_mask) && o_mask->is_UP() ) | |
| 102 // Here we assume a trip through memory is required. | |
| 103 w_i_mask = &C->FIRST_STACK_mask(); | |
| 104 } | |
| 105 return new (C) MachSpillCopyNode( def, *w_i_mask, *w_o_mask ); | |
| 106 } | |
| 107 | |
| 108 //------------------------------insert_proj------------------------------------ | |
| 605 | 109 // Insert the spill at chosen location. Skip over any intervening Proj's or |
| 0 | 110 // Phis. Skip over a CatchNode and projs, inserting in the fall-through block |
| 111 // instead. Update high-pressure indices. Create a new live range. | |
| 112 void PhaseChaitin::insert_proj( Block *b, uint i, Node *spill, uint maxlrg ) { | |
| 113 // Skip intervening ProjNodes. Do not insert between a ProjNode and | |
| 114 // its definer. | |
| 115 while( i < b->_nodes.size() && | |
| 116 (b->_nodes[i]->is_Proj() || | |
|
566
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
550
diff
changeset
|
117 b->_nodes[i]->is_Phi() ) ) |
| 0 | 118 i++; |
| 119 | |
| 120 // Do not insert between a call and his Catch | |
| 121 if( b->_nodes[i]->is_Catch() ) { | |
| 122 // Put the instruction at the top of the fall-thru block. | |
| 123 // Find the fall-thru projection | |
| 124 while( 1 ) { | |
| 125 const CatchProjNode *cp = b->_nodes[++i]->as_CatchProj(); | |
| 126 if( cp->_con == CatchProjNode::fall_through_index ) | |
| 127 break; | |
| 128 } | |
| 129 int sidx = i - b->end_idx()-1; | |
| 130 b = b->_succs[sidx]; // Switch to successor block | |
| 131 i = 1; // Right at start of block | |
| 132 } | |
| 133 | |
| 134 b->_nodes.insert(i,spill); // Insert node in block | |
|
12023
d1034bd8cefc
8022284: Hide internal data structure in PhaseCFG
adlertz
parents:
10395
diff
changeset
|
135 _cfg.map_node_to_block(spill, b); // Update node->block mapping to reflect |
| 0 | 136 // Adjust the point where we go hi-pressure |
| 137 if( i <= b->_ihrp_index ) b->_ihrp_index++; | |
| 138 if( i <= b->_fhrp_index ) b->_fhrp_index++; | |
| 139 | |
| 140 // Assign a new Live Range Number to the SpillCopy and grow | |
| 141 // the node->live range mapping. | |
| 142 new_lrg(spill,maxlrg); | |
| 143 } | |
| 144 | |
| 145 //------------------------------split_DEF-------------------------------------- | |
| 605 | 146 // There are four categories of Split; UP/DOWN x DEF/USE |
| 0 | 147 // Only three of these really occur as DOWN/USE will always color |
| 148 // Any Split with a DEF cannot CISC-Spill now. Thus we need | |
| 149 // two helper routines, one for Split DEFS (insert after instruction), | |
| 150 // one for Split USES (insert before instruction). DEF insertion | |
| 151 // happens inside Split, where the Leaveblock array is updated. | |
| 152 uint PhaseChaitin::split_DEF( Node *def, Block *b, int loc, uint maxlrg, Node **Reachblock, Node **debug_defs, GrowableArray<uint> splits, int slidx ) { | |
| 153 #ifdef ASSERT | |
| 154 // Increment the counter for this lrg | |
| 155 splits.at_put(slidx, splits.at(slidx)+1); | |
| 156 #endif | |
| 157 // If we are spilling the memory op for an implicit null check, at the | |
| 158 // null check location (ie - null check is in HRP block) we need to do | |
| 159 // the null-check first, then spill-down in the following block. | |
| 160 // (The implicit_null_check function ensures the use is also dominated | |
| 161 // by the branch-not-taken block.) | |
| 162 Node *be = b->end(); | |
| 163 if( be->is_MachNullCheck() && be->in(1) == def && def == b->_nodes[loc] ) { | |
| 164 // Spill goes in the branch-not-taken block | |
| 165 b = b->_succs[b->_nodes[b->end_idx()+1]->Opcode() == Op_IfTrue]; | |
| 166 loc = 0; // Just past the Region | |
| 167 } | |
| 168 assert( loc >= 0, "must insert past block head" ); | |
| 169 | |
| 170 // Get a def-side SpillCopy | |
| 171 Node *spill = get_spillcopy_wide(def,NULL,0); | |
| 172 // Did we fail to split?, then bail | |
| 173 if (!spill) { | |
| 174 return 0; | |
| 175 } | |
| 176 | |
| 177 // Insert the spill at chosen location | |
| 178 insert_proj( b, loc+1, spill, maxlrg++); | |
| 179 | |
| 180 // Insert new node into Reaches array | |
| 181 Reachblock[slidx] = spill; | |
| 182 // Update debug list of reaching down definitions by adding this one | |
| 183 debug_defs[slidx] = spill; | |
| 184 | |
| 185 // return updated count of live ranges | |
| 186 return maxlrg; | |
| 187 } | |
| 188 | |
| 189 //------------------------------split_USE-------------------------------------- | |
| 190 // Splits at uses can involve redeffing the LRG, so no CISC Spilling there. | |
| 191 // Debug uses want to know if def is already stack enabled. | |
| 192 uint PhaseChaitin::split_USE( Node *def, Block *b, Node *use, uint useidx, uint maxlrg, bool def_down, bool cisc_sp, GrowableArray<uint> splits, int slidx ) { | |
| 193 #ifdef ASSERT | |
| 194 // Increment the counter for this lrg | |
| 195 splits.at_put(slidx, splits.at(slidx)+1); | |
| 196 #endif | |
| 197 | |
| 198 // Some setup stuff for handling debug node uses | |
| 199 JVMState* jvms = use->jvms(); | |
| 200 uint debug_start = jvms ? jvms->debug_start() : 999999; | |
| 201 uint debug_end = jvms ? jvms->debug_end() : 999999; | |
| 202 | |
| 203 //------------------------------------------- | |
| 204 // Check for use of debug info | |
| 205 if (useidx >= debug_start && useidx < debug_end) { | |
| 206 // Actually it's perfectly legal for constant debug info to appear | |
| 207 // just unlikely. In this case the optimizer left a ConI of a 4 | |
| 208 // as both inputs to a Phi with only a debug use. It's a single-def | |
| 209 // live range of a rematerializable value. The live range spills, | |
| 210 // rematerializes and now the ConI directly feeds into the debug info. | |
| 211 // assert(!def->is_Con(), "constant debug info already constructed directly"); | |
| 212 | |
| 213 // Special split handling for Debug Info | |
| 214 // If DEF is DOWN, just hook the edge and return | |
| 215 // If DEF is UP, Split it DOWN for this USE. | |
| 216 if( def->is_Mach() ) { | |
| 217 if( def_down ) { | |
| 218 // DEF is DOWN, so connect USE directly to the DEF | |
| 219 use->set_req(useidx, def); | |
| 220 } else { | |
| 221 // Block and index where the use occurs. | |
|
12023
d1034bd8cefc
8022284: Hide internal data structure in PhaseCFG
adlertz
parents:
10395
diff
changeset
|
222 Block *b = _cfg.get_block_for_node(use); |
| 0 | 223 // Put the clone just prior to use |
| 224 int bindex = b->find_node(use); | |
| 225 // DEF is UP, so must copy it DOWN and hook in USE | |
| 226 // Insert SpillCopy before the USE, which uses DEF as its input, | |
| 227 // and defs a new live range, which is used by this node. | |
| 228 Node *spill = get_spillcopy_wide(def,use,useidx); | |
| 229 // did we fail to split? | |
| 230 if (!spill) { | |
| 231 // Bail | |
| 232 return 0; | |
| 233 } | |
| 234 // insert into basic block | |
| 235 insert_proj( b, bindex, spill, maxlrg++ ); | |
| 236 // Use the new split | |
| 237 use->set_req(useidx,spill); | |
| 238 } | |
| 239 // No further split handling needed for this use | |
| 240 return maxlrg; | |
| 241 } // End special splitting for debug info live range | |
| 242 } // If debug info | |
| 243 | |
| 244 // CISC-SPILLING | |
| 245 // Finally, check to see if USE is CISC-Spillable, and if so, | |
| 246 // gather_lrg_masks will add the flags bit to its mask, and | |
| 247 // no use side copy is needed. This frees up the live range | |
| 248 // register choices without causing copy coalescing, etc. | |
| 249 if( UseCISCSpill && cisc_sp ) { | |
| 250 int inp = use->cisc_operand(); | |
| 251 if( inp != AdlcVMDeps::Not_cisc_spillable ) | |
| 252 // Convert operand number to edge index number | |
| 253 inp = use->as_Mach()->operand_index(inp); | |
| 254 if( inp == (int)useidx ) { | |
| 255 use->set_req(useidx, def); | |
| 256 #ifndef PRODUCT | |
| 257 if( TraceCISCSpill ) { | |
| 258 tty->print(" set_split: "); | |
| 259 use->dump(); | |
| 260 } | |
| 261 #endif | |
| 262 return maxlrg; | |
| 263 } | |
| 264 } | |
| 265 | |
| 266 //------------------------------------------- | |
| 267 // Insert a Copy before the use | |
| 268 | |
| 269 // Block and index where the use occurs. | |
| 270 int bindex; | |
| 271 // Phi input spill-copys belong at the end of the prior block | |
| 272 if( use->is_Phi() ) { | |
|
12023
d1034bd8cefc
8022284: Hide internal data structure in PhaseCFG
adlertz
parents:
10395
diff
changeset
|
273 b = _cfg.get_block_for_node(b->pred(useidx)); |
| 0 | 274 bindex = b->end_idx(); |
| 275 } else { | |
| 276 // Put the clone just prior to use | |
| 277 bindex = b->find_node(use); | |
| 278 } | |
| 279 | |
| 280 Node *spill = get_spillcopy_wide( def, use, useidx ); | |
| 281 if( !spill ) return 0; // Bailed out | |
| 282 // Insert SpillCopy before the USE, which uses the reaching DEF as | |
| 283 // its input, and defs a new live range, which is used by this node. | |
| 284 insert_proj( b, bindex, spill, maxlrg++ ); | |
| 285 // Use the spill/clone | |
| 286 use->set_req(useidx,spill); | |
| 287 | |
| 288 // return updated live range count | |
| 289 return maxlrg; | |
| 290 } | |
| 291 | |
|
1693
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
292 //------------------------------clone_node---------------------------- |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
293 // Clone node with anti dependence check. |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
294 Node* clone_node(Node* def, Block *b, Compile* C) { |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
295 if (def->needs_anti_dependence_check()) { |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
296 #ifdef ASSERT |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
297 if (Verbose) { |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
298 tty->print_cr("RA attempts to clone node with anti_dependence:"); |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
299 def->dump(-1); tty->cr(); |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
300 tty->print_cr("into block:"); |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
301 b->dump(); |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
302 } |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
303 #endif |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
304 if (C->subsume_loads() == true && !C->failing()) { |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
305 // Retry with subsume_loads == false |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
306 // If this is the first failure, the sentinel string will "stick" |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
307 // to the Compile object, and the C2Compiler will see it and retry. |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
308 C->record_failure(C2Compiler::retry_no_subsuming_loads()); |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
309 } else { |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
310 // Bailout without retry |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
311 C->record_method_not_compilable("RA Split failed: attempt to clone node with anti_dependence"); |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
312 } |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
313 return 0; |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
314 } |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
315 return def->clone(); |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
316 } |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
317 |
| 0 | 318 //------------------------------split_Rematerialize---------------------------- |
| 319 // Clone a local copy of the def. | |
| 320 Node *PhaseChaitin::split_Rematerialize( Node *def, Block *b, uint insidx, uint &maxlrg, GrowableArray<uint> splits, int slidx, uint *lrg2reach, Node **Reachblock, bool walkThru ) { | |
| 321 // The input live ranges will be stretched to the site of the new | |
| 322 // instruction. They might be stretched past a def and will thus | |
| 323 // have the old and new values of the same live range alive at the | |
| 324 // same time - a definite no-no. Split out private copies of | |
| 325 // the inputs. | |
| 326 if( def->req() > 1 ) { | |
| 327 for( uint i = 1; i < def->req(); i++ ) { | |
| 328 Node *in = def->in(i); | |
| 329 // Check for single-def (LRG cannot redefined) | |
| 10111 | 330 uint lidx = _lrg_map.live_range_id(in); |
| 331 if (lidx >= _lrg_map.max_lrg_id()) { | |
| 332 continue; // Value is a recent spill-copy | |
| 333 } | |
| 334 if (lrgs(lidx).is_singledef()) { | |
| 335 continue; | |
| 336 } | |
| 0 | 337 |
|
12023
d1034bd8cefc
8022284: Hide internal data structure in PhaseCFG
adlertz
parents:
10395
diff
changeset
|
338 Block *b_def = _cfg.get_block_for_node(def); |
| 0 | 339 int idx_def = b_def->find_node(def); |
| 340 Node *in_spill = get_spillcopy_wide( in, def, i ); | |
| 341 if( !in_spill ) return 0; // Bailed out | |
| 342 insert_proj(b_def,idx_def,in_spill,maxlrg++); | |
| 343 if( b_def == b ) | |
| 344 insidx++; | |
| 345 def->set_req(i,in_spill); | |
| 346 } | |
| 347 } | |
| 348 | |
|
1693
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
349 Node *spill = clone_node(def, b, C); |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
350 if (spill == NULL || C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { |
| 0 | 351 // Check when generating nodes |
| 352 return 0; | |
| 353 } | |
| 354 | |
| 355 // See if any inputs are currently being spilled, and take the | |
| 356 // latest copy of spilled inputs. | |
| 357 if( spill->req() > 1 ) { | |
| 358 for( uint i = 1; i < spill->req(); i++ ) { | |
| 359 Node *in = spill->in(i); | |
| 10111 | 360 uint lidx = _lrg_map.find_id(in); |
| 0 | 361 |
| 362 // Walk backwards thru spill copy node intermediates | |
|
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
363 if (walkThru) { |
| 10111 | 364 while (in->is_SpillCopy() && lidx >= _lrg_map.max_lrg_id()) { |
| 0 | 365 in = in->in(1); |
| 10111 | 366 lidx = _lrg_map.find_id(in); |
| 0 | 367 } |
| 368 | |
| 10111 | 369 if (lidx < _lrg_map.max_lrg_id() && lrgs(lidx).is_multidef()) { |
|
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
370 // walkThru found a multidef LRG, which is unsafe to use, so |
|
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
371 // just keep the original def used in the clone. |
|
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
372 in = spill->in(i); |
| 10111 | 373 lidx = _lrg_map.find_id(in); |
|
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
374 } |
|
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
375 } |
|
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
376 |
| 10111 | 377 if (lidx < _lrg_map.max_lrg_id() && lrgs(lidx).reg() >= LRG::SPILL_REG) { |
| 0 | 378 Node *rdef = Reachblock[lrg2reach[lidx]]; |
| 10111 | 379 if (rdef) { |
| 380 spill->set_req(i, rdef); | |
| 381 } | |
| 0 | 382 } |
| 383 } | |
| 384 } | |
| 385 | |
| 386 | |
| 387 assert( spill->out_RegMask().is_UP(), "rematerialize to a reg" ); | |
| 388 // Rematerialized op is def->spilled+1 | |
| 389 set_was_spilled(spill); | |
| 390 if( _spilled_once.test(def->_idx) ) | |
| 391 set_was_spilled(spill); | |
| 392 | |
| 393 insert_proj( b, insidx, spill, maxlrg++ ); | |
| 394 #ifdef ASSERT | |
| 395 // Increment the counter for this lrg | |
| 396 splits.at_put(slidx, splits.at(slidx)+1); | |
| 397 #endif | |
| 398 // See if the cloned def kills any flags, and copy those kills as well | |
| 399 uint i = insidx+1; | |
| 10111 | 400 if( clone_projs( b, i, def, spill, maxlrg) ) { |
| 0 | 401 // Adjust the point where we go hi-pressure |
| 402 if( i <= b->_ihrp_index ) b->_ihrp_index++; | |
| 403 if( i <= b->_fhrp_index ) b->_fhrp_index++; | |
| 404 } | |
| 405 | |
| 406 return spill; | |
| 407 } | |
| 408 | |
| 409 //------------------------------is_high_pressure------------------------------- | |
| 410 // Function to compute whether or not this live range is "high pressure" | |
| 411 // in this block - whether it spills eagerly or not. | |
| 412 bool PhaseChaitin::is_high_pressure( Block *b, LRG *lrg, uint insidx ) { | |
| 413 if( lrg->_was_spilled1 ) return true; | |
| 414 // Forced spilling due to conflict? Then split only at binding uses | |
| 415 // or defs, not for supposed capacity problems. | |
| 416 // CNC - Turned off 7/8/99, causes too much spilling | |
| 417 // if( lrg->_is_bound ) return false; | |
| 418 | |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
419 // Use float pressure numbers for vectors. |
|
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
420 bool is_float_or_vector = lrg->_is_float || lrg->_is_vector; |
| 0 | 421 // Not yet reached the high-pressure cutoff point, so low pressure |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
422 uint hrp_idx = is_float_or_vector ? b->_fhrp_index : b->_ihrp_index; |
| 0 | 423 if( insidx < hrp_idx ) return false; |
| 424 // Register pressure for the block as a whole depends on reg class | |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
425 int block_pres = is_float_or_vector ? b->_freg_pressure : b->_reg_pressure; |
| 0 | 426 // Bound live ranges will split at the binding points first; |
| 427 // Intermediate splits should assume the live range's register set | |
| 428 // got "freed up" and that num_regs will become INT_PRESSURE. | |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
429 int bound_pres = is_float_or_vector ? FLOATPRESSURE : INTPRESSURE; |
| 0 | 430 // Effective register pressure limit. |
| 431 int lrg_pres = (lrg->get_invalid_mask_size() > lrg->num_regs()) | |
| 432 ? (lrg->get_invalid_mask_size() >> (lrg->num_regs()-1)) : bound_pres; | |
| 433 // High pressure if block pressure requires more register freedom | |
| 434 // than live range has. | |
| 435 return block_pres >= lrg_pres; | |
| 436 } | |
| 437 | |
| 438 | |
| 439 //------------------------------prompt_use--------------------------------- | |
| 440 // True if lidx is used before any real register is def'd in the block | |
| 441 bool PhaseChaitin::prompt_use( Block *b, uint lidx ) { | |
| 10111 | 442 if (lrgs(lidx)._was_spilled2) { |
| 443 return false; | |
| 444 } | |
| 0 | 445 |
| 446 // Scan block for 1st use. | |
| 447 for( uint i = 1; i <= b->end_idx(); i++ ) { | |
| 448 Node *n = b->_nodes[i]; | |
| 449 // Ignore PHI use, these can be up or down | |
| 10111 | 450 if (n->is_Phi()) { |
| 451 continue; | |
| 452 } | |
| 453 for (uint j = 1; j < n->req(); j++) { | |
| 454 if (_lrg_map.find_id(n->in(j)) == lidx) { | |
| 0 | 455 return true; // Found 1st use! |
| 10111 | 456 } |
| 457 } | |
| 458 if (n->out_RegMask().is_NotEmpty()) { | |
| 459 return false; | |
| 460 } | |
| 0 | 461 } |
| 462 return false; | |
| 463 } | |
| 464 | |
| 465 //------------------------------Split-------------------------------------- | |
| 466 //----------Split Routine---------- | |
| 467 // ***** NEW SPLITTING HEURISTIC ***** | |
| 468 // DEFS: If the DEF is in a High Register Pressure(HRP) Block, split there. | |
| 469 // Else, no split unless there is a HRP block between a DEF and | |
| 470 // one of its uses, and then split at the HRP block. | |
| 471 // | |
| 472 // USES: If USE is in HRP, split at use to leave main LRG on stack. | |
| 473 // Else, hoist LRG back up to register only (ie - split is also DEF) | |
| 474 // We will compute a new maxlrg as we go | |
| 6632 | 475 uint PhaseChaitin::Split(uint maxlrg, ResourceArea* split_arena) { |
| 0 | 476 NOT_PRODUCT( Compile::TracePhase t3("regAllocSplit", &_t_regAllocSplit, TimeCompiler); ) |
| 477 | |
| 6632 | 478 // Free thread local resources used by this method on exit. |
| 479 ResourceMark rm(split_arena); | |
| 480 | |
| 0 | 481 uint bidx, pidx, slidx, insidx, inpidx, twoidx; |
| 482 uint non_phi = 1, spill_cnt = 0; | |
| 483 Node **Reachblock; | |
| 484 Node *n1, *n2, *n3; | |
| 485 Node_List *defs,*phis; | |
| 486 bool *UPblock; | |
| 487 bool u1, u2, u3; | |
| 488 Block *b, *pred; | |
| 489 PhiNode *phi; | |
| 10111 | 490 GrowableArray<uint> lidxs(split_arena, maxlrg, 0, 0); |
| 0 | 491 |
| 492 // Array of counters to count splits per live range | |
| 10111 | 493 GrowableArray<uint> splits(split_arena, maxlrg, 0, 0); |
| 6632 | 494 |
| 495 #define NEW_SPLIT_ARRAY(type, size)\ | |
| 496 (type*) split_arena->allocate_bytes((size) * sizeof(type)) | |
| 0 | 497 |
| 498 //----------Setup Code---------- | |
| 499 // Create a convenient mapping from lrg numbers to reaches/leaves indices | |
| 10111 | 500 uint *lrg2reach = NEW_SPLIT_ARRAY(uint, maxlrg); |
| 0 | 501 // Keep track of DEFS & Phis for later passes |
| 502 defs = new Node_List(); | |
| 503 phis = new Node_List(); | |
| 504 // Gather info on which LRG's are spilling, and build maps | |
| 10111 | 505 for (bidx = 1; bidx < maxlrg; bidx++) { |
| 506 if (lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG) { | |
| 0 | 507 assert(!lrgs(bidx).mask().is_AllStack(),"AllStack should color"); |
| 508 lrg2reach[bidx] = spill_cnt; | |
| 509 spill_cnt++; | |
| 510 lidxs.append(bidx); | |
| 511 #ifdef ASSERT | |
| 512 // Initialize the split counts to zero | |
| 513 splits.append(0); | |
| 514 #endif | |
| 515 #ifndef PRODUCT | |
| 516 if( PrintOpto && WizardMode && lrgs(bidx)._was_spilled1 ) | |
| 517 tty->print_cr("Warning, 2nd spill of L%d",bidx); | |
| 518 #endif | |
| 519 } | |
| 520 } | |
| 521 | |
| 522 // Create side arrays for propagating reaching defs info. | |
| 523 // Each block needs a node pointer for each spilling live range for the | |
| 524 // Def which is live into the block. Phi nodes handle multiple input | |
| 525 // Defs by querying the output of their predecessor blocks and resolving | |
| 526 // them to a single Def at the phi. The pointer is updated for each | |
| 527 // Def in the block, and then becomes the output for the block when | |
| 528 // processing of the block is complete. We also need to track whether | |
| 529 // a Def is UP or DOWN. UP means that it should get a register (ie - | |
| 530 // it is always in LRP regions), and DOWN means that it is probably | |
| 531 // on the stack (ie - it crosses HRP regions). | |
| 6632 | 532 Node ***Reaches = NEW_SPLIT_ARRAY( Node**, _cfg._num_blocks+1 ); |
| 533 bool **UP = NEW_SPLIT_ARRAY( bool*, _cfg._num_blocks+1 ); | |
| 534 Node **debug_defs = NEW_SPLIT_ARRAY( Node*, spill_cnt ); | |
| 535 VectorSet **UP_entry= NEW_SPLIT_ARRAY( VectorSet*, spill_cnt ); | |
| 0 | 536 |
| 537 // Initialize Reaches & UP | |
| 538 for( bidx = 0; bidx < _cfg._num_blocks+1; bidx++ ) { | |
| 6632 | 539 Reaches[bidx] = NEW_SPLIT_ARRAY( Node*, spill_cnt ); |
| 540 UP[bidx] = NEW_SPLIT_ARRAY( bool, spill_cnt ); | |
| 0 | 541 Node **Reachblock = Reaches[bidx]; |
| 542 bool *UPblock = UP[bidx]; | |
| 543 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
| 544 UPblock[slidx] = true; // Assume they start in registers | |
| 545 Reachblock[slidx] = NULL; // Assume that no def is present | |
| 546 } | |
| 547 } | |
| 548 | |
| 6632 | 549 #undef NEW_SPLIT_ARRAY |
| 550 | |
| 0 | 551 // Initialize to array of empty vectorsets |
| 552 for( slidx = 0; slidx < spill_cnt; slidx++ ) | |
| 6632 | 553 UP_entry[slidx] = new VectorSet(split_arena); |
| 0 | 554 |
| 555 //----------PASS 1---------- | |
| 556 //----------Propagation & Node Insertion Code---------- | |
| 557 // Walk the Blocks in RPO for DEF & USE info | |
| 558 for( bidx = 0; bidx < _cfg._num_blocks; bidx++ ) { | |
| 559 | |
| 560 if (C->check_node_count(spill_cnt, out_of_nodes)) { | |
| 561 return 0; | |
| 562 } | |
| 563 | |
| 564 b = _cfg._blocks[bidx]; | |
| 565 // Reaches & UP arrays for this block | |
| 566 Reachblock = Reaches[b->_pre_order]; | |
| 567 UPblock = UP[b->_pre_order]; | |
| 568 // Reset counter of start of non-Phi nodes in block | |
| 569 non_phi = 1; | |
| 570 //----------Block Entry Handling---------- | |
| 571 // Check for need to insert a new phi | |
| 572 // Cycle through this block's predecessors, collecting Reaches | |
| 573 // info for each spilled LRG. If they are identical, no phi is | |
| 574 // needed. If they differ, check for a phi, and insert if missing, | |
| 575 // or update edges if present. Set current block's Reaches set to | |
| 576 // be either the phi's or the reaching def, as appropriate. | |
| 577 // If no Phi is needed, check if the LRG needs to spill on entry | |
| 578 // to the block due to HRP. | |
| 579 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
| 580 // Grab the live range number | |
| 581 uint lidx = lidxs.at(slidx); | |
| 582 // Do not bother splitting or putting in Phis for single-def | |
| 583 // rematerialized live ranges. This happens alot to constants | |
| 584 // with long live ranges. | |
|
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
585 if( lrgs(lidx).is_singledef() && |
| 0 | 586 lrgs(lidx)._def->rematerialize() ) { |
| 587 // reset the Reaches & UP entries | |
| 588 Reachblock[slidx] = lrgs(lidx)._def; | |
| 589 UPblock[slidx] = true; | |
| 590 // Record following instruction in case 'n' rematerializes and | |
| 591 // kills flags | |
|
12023
d1034bd8cefc
8022284: Hide internal data structure in PhaseCFG
adlertz
parents:
10395
diff
changeset
|
592 Block *pred1 = _cfg.get_block_for_node(b->pred(1)); |
| 0 | 593 continue; |
| 594 } | |
| 595 | |
| 596 // Initialize needs_phi and needs_split | |
| 597 bool needs_phi = false; | |
| 598 bool needs_split = false; | |
|
330
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
599 bool has_phi = false; |
| 0 | 600 // Walk the predecessor blocks to check inputs for that live range |
| 601 // Grab predecessor block header | |
| 602 n1 = b->pred(1); | |
| 603 // Grab the appropriate reaching def info for inpidx | |
|
12023
d1034bd8cefc
8022284: Hide internal data structure in PhaseCFG
adlertz
parents:
10395
diff
changeset
|
604 pred = _cfg.get_block_for_node(n1); |
| 0 | 605 pidx = pred->_pre_order; |
| 606 Node **Ltmp = Reaches[pidx]; | |
| 607 bool *Utmp = UP[pidx]; | |
| 608 n1 = Ltmp[slidx]; | |
| 609 u1 = Utmp[slidx]; | |
| 610 // Initialize node for saving type info | |
| 611 n3 = n1; | |
| 612 u3 = u1; | |
| 613 | |
| 614 // Compare inputs to see if a Phi is needed | |
| 615 for( inpidx = 2; inpidx < b->num_preds(); inpidx++ ) { | |
| 616 // Grab predecessor block headers | |
| 617 n2 = b->pred(inpidx); | |
| 618 // Grab the appropriate reaching def info for inpidx | |
|
12023
d1034bd8cefc
8022284: Hide internal data structure in PhaseCFG
adlertz
parents:
10395
diff
changeset
|
619 pred = _cfg.get_block_for_node(n2); |
| 0 | 620 pidx = pred->_pre_order; |
| 621 Ltmp = Reaches[pidx]; | |
| 622 Utmp = UP[pidx]; | |
| 623 n2 = Ltmp[slidx]; | |
| 624 u2 = Utmp[slidx]; | |
| 625 // For each LRG, decide if a phi is necessary | |
| 626 if( n1 != n2 ) { | |
| 627 needs_phi = true; | |
| 628 } | |
| 629 // See if the phi has mismatched inputs, UP vs. DOWN | |
| 630 if( n1 && n2 && (u1 != u2) ) { | |
| 631 needs_split = true; | |
| 632 } | |
| 633 // Move n2/u2 to n1/u1 for next iteration | |
| 634 n1 = n2; | |
| 635 u1 = u2; | |
| 636 // Preserve a non-NULL predecessor for later type referencing | |
| 637 if( (n3 == NULL) && (n2 != NULL) ){ | |
| 638 n3 = n2; | |
| 639 u3 = u2; | |
| 640 } | |
| 641 } // End for all potential Phi inputs | |
| 642 | |
|
330
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
643 // check block for appropriate phinode & update edges |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
644 for( insidx = 1; insidx <= b->end_idx(); insidx++ ) { |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
645 n1 = b->_nodes[insidx]; |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
646 // bail if this is not a phi |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
647 phi = n1->is_Phi() ? n1->as_Phi() : NULL; |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
648 if( phi == NULL ) { |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
649 // Keep track of index of first non-PhiNode instruction in block |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
650 non_phi = insidx; |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
651 // break out of the for loop as we have handled all phi nodes |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
652 break; |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
653 } |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
654 // must be looking at a phi |
| 10111 | 655 if (_lrg_map.find_id(n1) == lidxs.at(slidx)) { |
|
330
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
656 // found the necessary phi |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
657 needs_phi = false; |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
658 has_phi = true; |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
659 // initialize the Reaches entry for this LRG |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
660 Reachblock[slidx] = phi; |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
661 break; |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
662 } // end if found correct phi |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
663 } // end for all phi's |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
664 |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
665 // If a phi is needed or exist, check for it |
|
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
666 if( needs_phi || has_phi ) { |
| 0 | 667 // add new phinode if one not already found |
| 668 if( needs_phi ) { | |
| 669 // create a new phi node and insert it into the block | |
| 670 // type is taken from left over pointer to a predecessor | |
| 671 assert(n3,"No non-NULL reaching DEF for a Phi"); | |
|
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6725
diff
changeset
|
672 phi = new (C) PhiNode(b->head(), n3->bottom_type()); |
| 0 | 673 // initialize the Reaches entry for this LRG |
| 674 Reachblock[slidx] = phi; | |
| 675 | |
| 676 // add node to block & node_to_block mapping | |
| 10111 | 677 insert_proj(b, insidx++, phi, maxlrg++); |
| 0 | 678 non_phi++; |
| 679 // Reset new phi's mapping to be the spilling live range | |
| 10111 | 680 _lrg_map.map(phi->_idx, lidx); |
| 681 assert(_lrg_map.find_id(phi) == lidx, "Bad update on Union-Find mapping"); | |
| 0 | 682 } // end if not found correct phi |
| 683 // Here you have either found or created the Phi, so record it | |
| 684 assert(phi != NULL,"Must have a Phi Node here"); | |
| 685 phis->push(phi); | |
| 686 // PhiNodes should either force the LRG UP or DOWN depending | |
| 687 // on its inputs and the register pressure in the Phi's block. | |
| 688 UPblock[slidx] = true; // Assume new DEF is UP | |
| 689 // If entering a high-pressure area with no immediate use, | |
| 690 // assume Phi is DOWN | |
| 691 if( is_high_pressure( b, &lrgs(lidx), b->end_idx()) && !prompt_use(b,lidx) ) | |
| 692 UPblock[slidx] = false; | |
| 693 // If we are not split up/down and all inputs are down, then we | |
| 694 // are down | |
| 695 if( !needs_split && !u3 ) | |
| 696 UPblock[slidx] = false; | |
| 697 } // end if phi is needed | |
| 698 | |
| 699 // Do not need a phi, so grab the reaching DEF | |
| 700 else { | |
| 701 // Grab predecessor block header | |
| 702 n1 = b->pred(1); | |
| 703 // Grab the appropriate reaching def info for k | |
|
12023
d1034bd8cefc
8022284: Hide internal data structure in PhaseCFG
adlertz
parents:
10395
diff
changeset
|
704 pred = _cfg.get_block_for_node(n1); |
| 0 | 705 pidx = pred->_pre_order; |
| 706 Node **Ltmp = Reaches[pidx]; | |
| 707 bool *Utmp = UP[pidx]; | |
| 708 // reset the Reaches & UP entries | |
| 709 Reachblock[slidx] = Ltmp[slidx]; | |
| 710 UPblock[slidx] = Utmp[slidx]; | |
| 711 } // end else no Phi is needed | |
| 712 } // end for all spilling live ranges | |
| 713 // DEBUG | |
| 714 #ifndef PRODUCT | |
| 715 if(trace_spilling()) { | |
| 716 tty->print("/`\nBlock %d: ", b->_pre_order); | |
| 717 tty->print("Reaching Definitions after Phi handling\n"); | |
| 718 for( uint x = 0; x < spill_cnt; x++ ) { | |
| 719 tty->print("Spill Idx %d: UP %d: Node\n",x,UPblock[x]); | |
| 720 if( Reachblock[x] ) | |
| 721 Reachblock[x]->dump(); | |
| 722 else | |
| 723 tty->print("Undefined\n"); | |
| 724 } | |
| 725 } | |
| 726 #endif | |
| 727 | |
| 728 //----------Non-Phi Node Splitting---------- | |
| 729 // Since phi-nodes have now been handled, the Reachblock array for this | |
| 730 // block is initialized with the correct starting value for the defs which | |
| 731 // reach non-phi instructions in this block. Thus, process non-phi | |
| 732 // instructions normally, inserting SpillCopy nodes for all spill | |
| 733 // locations. | |
| 734 | |
| 735 // Memoize any DOWN reaching definitions for use as DEBUG info | |
| 736 for( insidx = 0; insidx < spill_cnt; insidx++ ) { | |
| 737 debug_defs[insidx] = (UPblock[insidx]) ? NULL : Reachblock[insidx]; | |
| 738 if( UPblock[insidx] ) // Memoize UP decision at block start | |
| 739 UP_entry[insidx]->set( b->_pre_order ); | |
| 740 } | |
| 741 | |
| 742 //----------Walk Instructions in the Block and Split---------- | |
| 743 // For all non-phi instructions in the block | |
| 744 for( insidx = 1; insidx <= b->end_idx(); insidx++ ) { | |
| 745 Node *n = b->_nodes[insidx]; | |
| 746 // Find the defining Node's live range index | |
| 10111 | 747 uint defidx = _lrg_map.find_id(n); |
| 0 | 748 uint cnt = n->req(); |
| 749 | |
| 10111 | 750 if (n->is_Phi()) { |
| 0 | 751 // Skip phi nodes after removing dead copies. |
| 10111 | 752 if (defidx < _lrg_map.max_lrg_id()) { |
| 0 | 753 // Check for useless Phis. These appear if we spill, then |
| 754 // coalesce away copies. Dont touch Phis in spilling live | |
| 755 // ranges; they are busy getting modifed in this pass. | |
| 756 if( lrgs(defidx).reg() < LRG::SPILL_REG ) { | |
| 757 uint i; | |
| 758 Node *u = NULL; | |
| 759 // Look for the Phi merging 2 unique inputs | |
| 760 for( i = 1; i < cnt; i++ ) { | |
| 761 // Ignore repeats and self | |
| 762 if( n->in(i) != u && n->in(i) != n ) { | |
| 763 // Found a unique input | |
| 764 if( u != NULL ) // If it's the 2nd, bail out | |
| 765 break; | |
| 766 u = n->in(i); // Else record it | |
| 767 } | |
| 768 } | |
| 769 assert( u, "at least 1 valid input expected" ); | |
| 10111 | 770 if (i >= cnt) { // Found one unique input |
| 771 assert(_lrg_map.find_id(n) == _lrg_map.find_id(u), "should be the same lrg"); | |
| 0 | 772 n->replace_by(u); // Then replace with unique input |
|
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6804
diff
changeset
|
773 n->disconnect_inputs(NULL, C); |
| 0 | 774 b->_nodes.remove(insidx); |
| 775 insidx--; | |
| 776 b->_ihrp_index--; | |
| 777 b->_fhrp_index--; | |
| 778 } | |
| 779 } | |
| 780 } | |
| 781 continue; | |
| 782 } | |
| 783 assert( insidx > b->_ihrp_index || | |
| 784 (b->_reg_pressure < (uint)INTPRESSURE) || | |
| 785 b->_ihrp_index > 4000000 || | |
| 786 b->_ihrp_index >= b->end_idx() || | |
| 787 !b->_nodes[b->_ihrp_index]->is_Proj(), "" ); | |
| 788 assert( insidx > b->_fhrp_index || | |
| 789 (b->_freg_pressure < (uint)FLOATPRESSURE) || | |
| 790 b->_fhrp_index > 4000000 || | |
| 791 b->_fhrp_index >= b->end_idx() || | |
| 792 !b->_nodes[b->_fhrp_index]->is_Proj(), "" ); | |
| 793 | |
| 794 // ********** Handle Crossing HRP Boundry ********** | |
| 795 if( (insidx == b->_ihrp_index) || (insidx == b->_fhrp_index) ) { | |
| 796 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
| 605 | 797 // Check for need to split at HRP boundary - split if UP |
| 0 | 798 n1 = Reachblock[slidx]; |
| 799 // bail out if no reaching DEF | |
| 800 if( n1 == NULL ) continue; | |
| 801 // bail out if live range is 'isolated' around inner loop | |
| 802 uint lidx = lidxs.at(slidx); | |
| 803 // If live range is currently UP | |
| 804 if( UPblock[slidx] ) { | |
| 805 // set location to insert spills at | |
| 806 // SPLIT DOWN HERE - NO CISC SPILL | |
| 807 if( is_high_pressure( b, &lrgs(lidx), insidx ) && | |
| 808 !n1->rematerialize() ) { | |
| 809 // If there is already a valid stack definition available, use it | |
| 810 if( debug_defs[slidx] != NULL ) { | |
| 811 Reachblock[slidx] = debug_defs[slidx]; | |
| 812 } | |
| 813 else { | |
| 814 // Insert point is just past last use or def in the block | |
| 815 int insert_point = insidx-1; | |
| 816 while( insert_point > 0 ) { | |
| 817 Node *n = b->_nodes[insert_point]; | |
| 818 // Hit top of block? Quit going backwards | |
| 10111 | 819 if (n->is_Phi()) { |
| 820 break; | |
| 821 } | |
| 0 | 822 // Found a def? Better split after it. |
| 10111 | 823 if (_lrg_map.live_range_id(n) == lidx) { |
| 824 break; | |
| 825 } | |
| 0 | 826 // Look for a use |
| 827 uint i; | |
| 10111 | 828 for( i = 1; i < n->req(); i++ ) { |
| 829 if (_lrg_map.live_range_id(n->in(i)) == lidx) { | |
| 0 | 830 break; |
| 10111 | 831 } |
| 832 } | |
| 0 | 833 // Found a use? Better split after it. |
| 10111 | 834 if (i < n->req()) { |
| 835 break; | |
| 836 } | |
| 0 | 837 insert_point--; |
| 838 } | |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
839 uint orig_eidx = b->end_idx(); |
| 0 | 840 maxlrg = split_DEF( n1, b, insert_point, maxlrg, Reachblock, debug_defs, splits, slidx); |
| 841 // If it wasn't split bail | |
| 842 if (!maxlrg) { | |
| 843 return 0; | |
| 844 } | |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
845 // Spill of NULL check mem op goes into the following block. |
| 10111 | 846 if (b->end_idx() > orig_eidx) { |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
847 insidx++; |
| 10111 | 848 } |
| 0 | 849 } |
| 850 // This is a new DEF, so update UP | |
| 851 UPblock[slidx] = false; | |
| 852 #ifndef PRODUCT | |
| 853 // DEBUG | |
| 854 if( trace_spilling() ) { | |
| 855 tty->print("\nNew Split DOWN DEF of Spill Idx "); | |
| 856 tty->print("%d, UP %d:\n",slidx,false); | |
| 857 n1->dump(); | |
| 858 } | |
| 859 #endif | |
| 860 } | |
| 861 } // end if LRG is UP | |
| 862 } // end for all spilling live ranges | |
| 863 assert( b->_nodes[insidx] == n, "got insidx set incorrectly" ); | |
| 864 } // end if crossing HRP Boundry | |
| 865 | |
| 866 // If the LRG index is oob, then this is a new spillcopy, skip it. | |
| 10111 | 867 if (defidx >= _lrg_map.max_lrg_id()) { |
| 0 | 868 continue; |
| 869 } | |
| 870 LRG &deflrg = lrgs(defidx); | |
| 871 uint copyidx = n->is_Copy(); | |
| 872 // Remove coalesced copy from CFG | |
| 10111 | 873 if (copyidx && defidx == _lrg_map.live_range_id(n->in(copyidx))) { |
| 0 | 874 n->replace_by( n->in(copyidx) ); |
| 875 n->set_req( copyidx, NULL ); | |
| 876 b->_nodes.remove(insidx--); | |
| 877 b->_ihrp_index--; // Adjust the point where we go hi-pressure | |
| 878 b->_fhrp_index--; | |
| 879 continue; | |
| 880 } | |
| 881 | |
| 882 #define DERIVED 0 | |
| 883 | |
| 884 // ********** Handle USES ********** | |
| 885 bool nullcheck = false; | |
| 886 // Implicit null checks never use the spilled value | |
| 887 if( n->is_MachNullCheck() ) | |
| 888 nullcheck = true; | |
| 889 if( !nullcheck ) { | |
| 890 // Search all inputs for a Spill-USE | |
| 891 JVMState* jvms = n->jvms(); | |
| 892 uint oopoff = jvms ? jvms->oopoff() : cnt; | |
| 893 uint old_last = cnt - 1; | |
| 894 for( inpidx = 1; inpidx < cnt; inpidx++ ) { | |
| 895 // Derived/base pairs may be added to our inputs during this loop. | |
| 896 // If inpidx > old_last, then one of these new inputs is being | |
| 897 // handled. Skip the derived part of the pair, but process | |
| 898 // the base like any other input. | |
| 10111 | 899 if (inpidx > old_last && ((inpidx - oopoff) & 1) == DERIVED) { |
| 0 | 900 continue; // skip derived_debug added below |
| 901 } | |
| 902 // Get lidx of input | |
| 10111 | 903 uint useidx = _lrg_map.find_id(n->in(inpidx)); |
| 0 | 904 // Not a brand-new split, and it is a spill use |
| 10111 | 905 if (useidx < _lrg_map.max_lrg_id() && lrgs(useidx).reg() >= LRG::SPILL_REG) { |
| 0 | 906 // Check for valid reaching DEF |
| 907 slidx = lrg2reach[useidx]; | |
| 908 Node *def = Reachblock[slidx]; | |
| 909 assert( def != NULL, "Using Undefined Value in Split()\n"); | |
| 910 | |
| 911 // (+++) %%%% remove this in favor of pre-pass in matcher.cpp | |
| 912 // monitor references do not care where they live, so just hook | |
| 913 if ( jvms && jvms->is_monitor_use(inpidx) ) { | |
| 914 // The effect of this clone is to drop the node out of the block, | |
| 915 // so that the allocator does not see it anymore, and therefore | |
| 916 // does not attempt to assign it a register. | |
|
1693
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
917 def = clone_node(def, b, C); |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
918 if (def == NULL || C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
919 return 0; |
|
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
920 } |
| 10111 | 921 _lrg_map.extend(def->_idx, 0); |
|
12023
d1034bd8cefc
8022284: Hide internal data structure in PhaseCFG
adlertz
parents:
10395
diff
changeset
|
922 _cfg.map_node_to_block(def, b); |
| 0 | 923 n->set_req(inpidx, def); |
| 924 continue; | |
| 925 } | |
| 926 | |
| 927 // Rematerializable? Then clone def at use site instead | |
| 928 // of store/load | |
| 929 if( def->rematerialize() ) { | |
| 930 int old_size = b->_nodes.size(); | |
| 931 def = split_Rematerialize( def, b, insidx, maxlrg, splits, slidx, lrg2reach, Reachblock, true ); | |
| 932 if( !def ) return 0; // Bail out | |
| 933 insidx += b->_nodes.size()-old_size; | |
| 934 } | |
| 935 | |
| 936 MachNode *mach = n->is_Mach() ? n->as_Mach() : NULL; | |
| 937 // Base pointers and oopmap references do not care where they live. | |
| 938 if ((inpidx >= oopoff) || | |
| 939 (mach && mach->ideal_Opcode() == Op_AddP && inpidx == AddPNode::Base)) { | |
| 940 if (def->rematerialize() && lrgs(useidx)._was_spilled2) { | |
| 941 // This def has been rematerialized a couple of times without | |
| 942 // progress. It doesn't care if it lives UP or DOWN, so | |
| 943 // spill it down now. | |
| 944 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false,splits,slidx); | |
| 945 // If it wasn't split bail | |
| 946 if (!maxlrg) { | |
| 947 return 0; | |
| 948 } | |
| 949 insidx++; // Reset iterator to skip USE side split | |
| 950 } else { | |
| 951 // Just hook the def edge | |
| 952 n->set_req(inpidx, def); | |
| 953 } | |
| 954 | |
| 955 if (inpidx >= oopoff) { | |
| 956 // After oopoff, we have derived/base pairs. We must mention all | |
| 957 // derived pointers here as derived/base pairs for GC. If the | |
| 958 // derived value is spilling and we have a copy both in Reachblock | |
| 959 // (called here 'def') and debug_defs[slidx] we need to mention | |
| 960 // both in derived/base pairs or kill one. | |
| 961 Node *derived_debug = debug_defs[slidx]; | |
| 962 if( ((inpidx - oopoff) & 1) == DERIVED && // derived vs base? | |
| 963 mach && mach->ideal_Opcode() != Op_Halt && | |
| 964 derived_debug != NULL && | |
| 965 derived_debug != def ) { // Actual 2nd value appears | |
| 966 // We have already set 'def' as a derived value. | |
| 967 // Also set debug_defs[slidx] as a derived value. | |
| 968 uint k; | |
| 969 for( k = oopoff; k < cnt; k += 2 ) | |
| 970 if( n->in(k) == derived_debug ) | |
| 971 break; // Found an instance of debug derived | |
| 972 if( k == cnt ) {// No instance of debug_defs[slidx] | |
| 973 // Add a derived/base pair to cover the debug info. | |
| 974 // We have to process the added base later since it is not | |
| 975 // handled yet at this point but skip derived part. | |
| 976 assert(((n->req() - oopoff) & 1) == DERIVED, | |
| 977 "must match skip condition above"); | |
| 978 n->add_req( derived_debug ); // this will be skipped above | |
| 979 n->add_req( n->in(inpidx+1) ); // this will be processed | |
| 980 // Increment cnt to handle added input edges on | |
| 981 // subsequent iterations. | |
| 982 cnt += 2; | |
| 983 } | |
| 984 } | |
| 985 } | |
| 986 continue; | |
| 987 } | |
| 988 // Special logic for DEBUG info | |
| 989 if( jvms && b->_freq > BLOCK_FREQUENCY(0.5) ) { | |
| 990 uint debug_start = jvms->debug_start(); | |
| 991 // If this is debug info use & there is a reaching DOWN def | |
| 992 if ((debug_start <= inpidx) && (debug_defs[slidx] != NULL)) { | |
| 993 assert(inpidx < oopoff, "handle only debug info here"); | |
| 994 // Just hook it in & move on | |
| 995 n->set_req(inpidx, debug_defs[slidx]); | |
| 996 // (Note that this can make two sides of a split live at the | |
| 997 // same time: The debug def on stack, and another def in a | |
| 998 // register. The GC needs to know about both of them, but any | |
| 999 // derived pointers after oopoff will refer to only one of the | |
| 1000 // two defs and the GC would therefore miss the other. Thus | |
| 1001 // this hack is only allowed for debug info which is Java state | |
| 1002 // and therefore never a derived pointer.) | |
| 1003 continue; | |
| 1004 } | |
| 1005 } | |
| 1006 // Grab register mask info | |
| 1007 const RegMask &dmask = def->out_RegMask(); | |
| 1008 const RegMask &umask = n->in_RegMask(inpidx); | |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
1009 bool is_vect = RegMask::is_vector(def->ideal_reg()); |
| 0 | 1010 assert(inpidx < oopoff, "cannot use-split oop map info"); |
| 1011 | |
| 1012 bool dup = UPblock[slidx]; | |
| 1013 bool uup = umask.is_UP(); | |
| 1014 | |
| 1015 // Need special logic to handle bound USES. Insert a split at this | |
| 1016 // bound use if we can't rematerialize the def, or if we need the | |
| 1017 // split to form a misaligned pair. | |
| 1018 if( !umask.is_AllStack() && | |
| 1019 (int)umask.Size() <= lrgs(useidx).num_regs() && | |
| 1020 (!def->rematerialize() || | |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
1021 !is_vect && umask.is_misaligned_pair())) { |
| 0 | 1022 // These need a Split regardless of overlap or pressure |
| 1023 // SPLIT - NO DEF - NO CISC SPILL | |
| 1024 maxlrg = split_USE(def,b,n,inpidx,maxlrg,dup,false, splits,slidx); | |
| 1025 // If it wasn't split bail | |
| 1026 if (!maxlrg) { | |
| 1027 return 0; | |
| 1028 } | |
| 1029 insidx++; // Reset iterator to skip USE side split | |
| 1030 continue; | |
| 1031 } | |
|
1730
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
1032 |
| 3842 | 1033 if (UseFPUForSpilling && n->is_MachCall() && !uup && !dup ) { |
|
1730
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
1034 // The use at the call can force the def down so insert |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
1035 // a split before the use to allow the def more freedom. |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
1036 maxlrg = split_USE(def,b,n,inpidx,maxlrg,dup,false, splits,slidx); |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
1037 // If it wasn't split bail |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
1038 if (!maxlrg) { |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
1039 return 0; |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
1040 } |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
1041 insidx++; // Reset iterator to skip USE side split |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
1042 continue; |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
1043 } |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
1044 |
| 0 | 1045 // Here is the logic chart which describes USE Splitting: |
| 1046 // 0 = false or DOWN, 1 = true or UP | |
| 1047 // | |
| 1048 // Overlap | DEF | USE | Action | |
| 1049 //------------------------------------------------------- | |
| 1050 // 0 | 0 | 0 | Copy - mem -> mem | |
| 1051 // 0 | 0 | 1 | Split-UP - Check HRP | |
| 1052 // 0 | 1 | 0 | Split-DOWN - Debug Info? | |
| 1053 // 0 | 1 | 1 | Copy - reg -> reg | |
| 1054 // 1 | 0 | 0 | Reset Input Edge (no Split) | |
| 1055 // 1 | 0 | 1 | Split-UP - Check HRP | |
| 1056 // 1 | 1 | 0 | Split-DOWN - Debug Info? | |
| 1057 // 1 | 1 | 1 | Reset Input Edge (no Split) | |
| 1058 // | |
| 1059 // So, if (dup == uup), then overlap test determines action, | |
| 1060 // with true being no split, and false being copy. Else, | |
| 1061 // if DEF is DOWN, Split-UP, and check HRP to decide on | |
| 1062 // resetting DEF. Finally if DEF is UP, Split-DOWN, with | |
| 1063 // special handling for Debug Info. | |
| 1064 if( dup == uup ) { | |
| 1065 if( dmask.overlap(umask) ) { | |
| 1066 // Both are either up or down, and there is overlap, No Split | |
| 1067 n->set_req(inpidx, def); | |
| 1068 } | |
| 1069 else { // Both are either up or down, and there is no overlap | |
| 1070 if( dup ) { // If UP, reg->reg copy | |
| 1071 // COPY ACROSS HERE - NO DEF - NO CISC SPILL | |
| 1072 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false, splits,slidx); | |
| 1073 // If it wasn't split bail | |
| 1074 if (!maxlrg) { | |
| 1075 return 0; | |
| 1076 } | |
| 1077 insidx++; // Reset iterator to skip USE side split | |
| 1078 } | |
| 1079 else { // DOWN, mem->mem copy | |
| 1080 // COPY UP & DOWN HERE - NO DEF - NO CISC SPILL | |
| 1081 // First Split-UP to move value into Register | |
| 1082 uint def_ideal = def->ideal_reg(); | |
| 1083 const RegMask* tmp_rm = Matcher::idealreg2regmask[def_ideal]; | |
| 1084 Node *spill = new (C) MachSpillCopyNode(def, dmask, *tmp_rm); | |
| 1085 insert_proj( b, insidx, spill, maxlrg ); | |
| 1086 // Then Split-DOWN as if previous Split was DEF | |
| 1087 maxlrg = split_USE(spill,b,n,inpidx,maxlrg,false,false, splits,slidx); | |
| 1088 // If it wasn't split bail | |
| 1089 if (!maxlrg) { | |
| 1090 return 0; | |
| 1091 } | |
| 1092 insidx += 2; // Reset iterator to skip USE side splits | |
| 1093 } | |
| 1094 } // End else no overlap | |
| 1095 } // End if dup == uup | |
| 1096 // dup != uup, so check dup for direction of Split | |
| 1097 else { | |
| 1098 if( dup ) { // If UP, Split-DOWN and check Debug Info | |
| 1099 // If this node is already a SpillCopy, just patch the edge | |
| 1100 // except the case of spilling to stack. | |
| 1101 if( n->is_SpillCopy() ) { | |
| 1102 RegMask tmp_rm(umask); | |
| 1103 tmp_rm.SUBTRACT(Matcher::STACK_ONLY_mask); | |
| 1104 if( dmask.overlap(tmp_rm) ) { | |
| 1105 if( def != n->in(inpidx) ) { | |
| 1106 n->set_req(inpidx, def); | |
| 1107 } | |
| 1108 continue; | |
| 1109 } | |
| 1110 } | |
| 1111 // COPY DOWN HERE - NO DEF - NO CISC SPILL | |
| 1112 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false, splits,slidx); | |
| 1113 // If it wasn't split bail | |
| 1114 if (!maxlrg) { | |
| 1115 return 0; | |
| 1116 } | |
| 1117 insidx++; // Reset iterator to skip USE side split | |
| 1118 // Check for debug-info split. Capture it for later | |
| 1119 // debug splits of the same value | |
| 1120 if (jvms && jvms->debug_start() <= inpidx && inpidx < oopoff) | |
| 1121 debug_defs[slidx] = n->in(inpidx); | |
| 1122 | |
| 1123 } | |
| 1124 else { // DOWN, Split-UP and check register pressure | |
| 1125 if( is_high_pressure( b, &lrgs(useidx), insidx ) ) { | |
| 1126 // COPY UP HERE - NO DEF - CISC SPILL | |
| 1127 maxlrg = split_USE(def,b,n,inpidx,maxlrg,true,true, splits,slidx); | |
| 1128 // If it wasn't split bail | |
| 1129 if (!maxlrg) { | |
| 1130 return 0; | |
| 1131 } | |
| 1132 insidx++; // Reset iterator to skip USE side split | |
| 1133 } else { // LRP | |
| 1134 // COPY UP HERE - WITH DEF - NO CISC SPILL | |
| 1135 maxlrg = split_USE(def,b,n,inpidx,maxlrg,true,false, splits,slidx); | |
| 1136 // If it wasn't split bail | |
| 1137 if (!maxlrg) { | |
| 1138 return 0; | |
| 1139 } | |
| 1140 // Flag this lift-up in a low-pressure block as | |
| 1141 // already-spilled, so if it spills again it will | |
| 1142 // spill hard (instead of not spilling hard and | |
| 1143 // coalescing away). | |
| 1144 set_was_spilled(n->in(inpidx)); | |
| 1145 // Since this is a new DEF, update Reachblock & UP | |
| 1146 Reachblock[slidx] = n->in(inpidx); | |
| 1147 UPblock[slidx] = true; | |
| 1148 insidx++; // Reset iterator to skip USE side split | |
| 1149 } | |
| 1150 } // End else DOWN | |
| 1151 } // End dup != uup | |
| 1152 } // End if Spill USE | |
| 1153 } // End For All Inputs | |
| 1154 } // End If not nullcheck | |
| 1155 | |
| 1156 // ********** Handle DEFS ********** | |
| 1157 // DEFS either Split DOWN in HRP regions or when the LRG is bound, or | |
| 1158 // just reset the Reaches info in LRP regions. DEFS must always update | |
| 1159 // UP info. | |
| 1160 if( deflrg.reg() >= LRG::SPILL_REG ) { // Spilled? | |
| 1161 uint slidx = lrg2reach[defidx]; | |
| 1162 // Add to defs list for later assignment of new live range number | |
| 1163 defs->push(n); | |
| 1164 // Set a flag on the Node indicating it has already spilled. | |
| 1165 // Only do it for capacity spills not conflict spills. | |
| 1166 if( !deflrg._direct_conflict ) | |
| 1167 set_was_spilled(n); | |
| 1168 assert(!n->is_Phi(),"Cannot insert Phi into DEFS list"); | |
| 1169 // Grab UP info for DEF | |
| 1170 const RegMask &dmask = n->out_RegMask(); | |
| 1171 bool defup = dmask.is_UP(); | |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
1172 int ireg = n->ideal_reg(); |
|
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
1173 bool is_vect = RegMask::is_vector(ireg); |
| 0 | 1174 // Only split at Def if this is a HRP block or bound (and spilled once) |
| 1175 if( !n->rematerialize() && | |
|
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
1176 (((dmask.is_bound(ireg) || !is_vect && dmask.is_misaligned_pair()) && |
|
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
3842
diff
changeset
|
1177 (deflrg._direct_conflict || deflrg._must_spill)) || |
| 0 | 1178 // Check for LRG being up in a register and we are inside a high |
| 1179 // pressure area. Spill it down immediately. | |
| 1180 (defup && is_high_pressure(b,&deflrg,insidx))) ) { | |
| 1181 assert( !n->rematerialize(), "" ); | |
| 1182 assert( !n->is_SpillCopy(), "" ); | |
| 1183 // Do a split at the def site. | |
| 1184 maxlrg = split_DEF( n, b, insidx, maxlrg, Reachblock, debug_defs, splits, slidx ); | |
| 1185 // If it wasn't split bail | |
| 1186 if (!maxlrg) { | |
| 1187 return 0; | |
| 1188 } | |
| 1189 // Split DEF's Down | |
| 1190 UPblock[slidx] = 0; | |
| 1191 #ifndef PRODUCT | |
| 1192 // DEBUG | |
| 1193 if( trace_spilling() ) { | |
| 1194 tty->print("\nNew Split DOWN DEF of Spill Idx "); | |
| 1195 tty->print("%d, UP %d:\n",slidx,false); | |
| 1196 n->dump(); | |
| 1197 } | |
| 1198 #endif | |
| 1199 } | |
| 1200 else { // Neither bound nor HRP, must be LRP | |
| 1201 // otherwise, just record the def | |
| 1202 Reachblock[slidx] = n; | |
| 1203 // UP should come from the outRegmask() of the DEF | |
| 1204 UPblock[slidx] = defup; | |
| 1205 // Update debug list of reaching down definitions, kill if DEF is UP | |
| 1206 debug_defs[slidx] = defup ? NULL : n; | |
| 1207 #ifndef PRODUCT | |
| 1208 // DEBUG | |
| 1209 if( trace_spilling() ) { | |
| 1210 tty->print("\nNew DEF of Spill Idx "); | |
| 1211 tty->print("%d, UP %d:\n",slidx,defup); | |
| 1212 n->dump(); | |
| 1213 } | |
| 1214 #endif | |
| 1215 } // End else LRP | |
| 1216 } // End if spill def | |
| 1217 | |
| 1218 // ********** Split Left Over Mem-Mem Moves ********** | |
| 1219 // Check for mem-mem copies and split them now. Do not do this | |
| 1220 // to copies about to be spilled; they will be Split shortly. | |
| 10111 | 1221 if (copyidx) { |
| 0 | 1222 Node *use = n->in(copyidx); |
| 10111 | 1223 uint useidx = _lrg_map.find_id(use); |
| 1224 if (useidx < _lrg_map.max_lrg_id() && // This is not a new split | |
| 0 | 1225 OptoReg::is_stack(deflrg.reg()) && |
| 1226 deflrg.reg() < LRG::SPILL_REG ) { // And DEF is from stack | |
| 1227 LRG &uselrg = lrgs(useidx); | |
| 1228 if( OptoReg::is_stack(uselrg.reg()) && | |
| 1229 uselrg.reg() < LRG::SPILL_REG && // USE is from stack | |
| 1230 deflrg.reg() != uselrg.reg() ) { // Not trivially removed | |
|
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6632
diff
changeset
|
1231 uint def_ideal_reg = n->bottom_type()->ideal_reg(); |
| 0 | 1232 const RegMask &def_rm = *Matcher::idealreg2regmask[def_ideal_reg]; |
| 1233 const RegMask &use_rm = n->in_RegMask(copyidx); | |
| 1234 if( def_rm.overlap(use_rm) && n->is_SpillCopy() ) { // Bug 4707800, 'n' may be a storeSSL | |
| 1235 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { // Check when generating nodes | |
| 1236 return 0; | |
| 1237 } | |
| 1238 Node *spill = new (C) MachSpillCopyNode(use,use_rm,def_rm); | |
| 1239 n->set_req(copyidx,spill); | |
| 1240 n->as_MachSpillCopy()->set_in_RegMask(def_rm); | |
| 1241 // Put the spill just before the copy | |
| 1242 insert_proj( b, insidx++, spill, maxlrg++ ); | |
| 1243 } | |
| 1244 } | |
| 1245 } | |
| 1246 } | |
| 1247 } // End For All Instructions in Block - Non-PHI Pass | |
| 1248 | |
| 1249 // Check if each LRG is live out of this block so as not to propagate | |
| 1250 // beyond the last use of a LRG. | |
| 1251 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
| 1252 uint defidx = lidxs.at(slidx); | |
| 1253 IndexSet *liveout = _live->live(b); | |
| 1254 if( !liveout->member(defidx) ) { | |
| 1255 #ifdef ASSERT | |
| 1256 // The index defidx is not live. Check the liveout array to ensure that | |
| 1257 // it contains no members which compress to defidx. Finding such an | |
| 1258 // instance may be a case to add liveout adjustment in compress_uf_map(). | |
| 1259 // See 5063219. | |
| 1260 uint member; | |
| 1261 IndexSetIterator isi(liveout); | |
| 1262 while ((member = isi.next()) != 0) { | |
| 10111 | 1263 assert(defidx != _lrg_map.find_const(member), "Live out member has not been compressed"); |
| 0 | 1264 } |
| 1265 #endif | |
| 1266 Reachblock[slidx] = NULL; | |
| 1267 } else { | |
| 1268 assert(Reachblock[slidx] != NULL,"No reaching definition for liveout value"); | |
| 1269 } | |
| 1270 } | |
| 1271 #ifndef PRODUCT | |
| 1272 if( trace_spilling() ) | |
| 1273 b->dump(); | |
| 1274 #endif | |
| 1275 } // End For All Blocks | |
| 1276 | |
| 1277 //----------PASS 2---------- | |
| 1278 // Reset all DEF live range numbers here | |
| 1279 for( insidx = 0; insidx < defs->size(); insidx++ ) { | |
| 1280 // Grab the def | |
| 1281 n1 = defs->at(insidx); | |
| 1282 // Set new lidx for DEF | |
| 1283 new_lrg(n1, maxlrg++); | |
| 1284 } | |
| 1285 //----------Phi Node Splitting---------- | |
| 1286 // Clean up a phi here, and assign a new live range number | |
| 1287 // Cycle through this block's predecessors, collecting Reaches | |
| 1288 // info for each spilled LRG and update edges. | |
| 1289 // Walk the phis list to patch inputs, split phis, and name phis | |
|
2016
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1290 uint lrgs_before_phi_split = maxlrg; |
| 0 | 1291 for( insidx = 0; insidx < phis->size(); insidx++ ) { |
| 1292 Node *phi = phis->at(insidx); | |
| 1293 assert(phi->is_Phi(),"This list must only contain Phi Nodes"); | |
|
12023
d1034bd8cefc
8022284: Hide internal data structure in PhaseCFG
adlertz
parents:
10395
diff
changeset
|
1294 Block *b = _cfg.get_block_for_node(phi); |
| 0 | 1295 // Grab the live range number |
| 10111 | 1296 uint lidx = _lrg_map.find_id(phi); |
| 0 | 1297 uint slidx = lrg2reach[lidx]; |
| 1298 // Update node to lidx map | |
| 1299 new_lrg(phi, maxlrg++); | |
| 1300 // Get PASS1's up/down decision for the block. | |
| 1301 int phi_up = !!UP_entry[slidx]->test(b->_pre_order); | |
| 1302 | |
| 1303 // Force down if double-spilling live range | |
| 1304 if( lrgs(lidx)._was_spilled1 ) | |
| 1305 phi_up = false; | |
| 1306 | |
| 1307 // When splitting a Phi we an split it normal or "inverted". | |
| 1308 // An inverted split makes the splits target the Phi's UP/DOWN | |
| 1309 // sense inverted; then the Phi is followed by a final def-side | |
| 1310 // split to invert back. It changes which blocks the spill code | |
| 1311 // goes in. | |
| 1312 | |
| 1313 // Walk the predecessor blocks and assign the reaching def to the Phi. | |
| 1314 // Split Phi nodes by placing USE side splits wherever the reaching | |
| 1315 // DEF has the wrong UP/DOWN value. | |
| 1316 for( uint i = 1; i < b->num_preds(); i++ ) { | |
| 1317 // Get predecessor block pre-order number | |
|
12023
d1034bd8cefc
8022284: Hide internal data structure in PhaseCFG
adlertz
parents:
10395
diff
changeset
|
1318 Block *pred = _cfg.get_block_for_node(b->pred(i)); |
| 0 | 1319 pidx = pred->_pre_order; |
| 1320 // Grab reaching def | |
| 1321 Node *def = Reaches[pidx][slidx]; | |
| 1322 assert( def, "must have reaching def" ); | |
| 1323 // If input up/down sense and reg-pressure DISagree | |
|
10395
b274ac1dbe11
8005956: C2: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG defined in this block
adlertz
parents:
10111
diff
changeset
|
1324 if (def->rematerialize() && contains_no_live_range_input(def)) { |
|
2016
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1325 // Place the rematerialized node above any MSCs created during |
|
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1326 // phi node splitting. end_idx points at the insertion point |
|
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1327 // so look at the node before it. |
|
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1328 int insert = pred->end_idx(); |
|
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1329 while (insert >= 1 && |
|
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1330 pred->_nodes[insert - 1]->is_SpillCopy() && |
| 10111 | 1331 _lrg_map.find(pred->_nodes[insert - 1]) >= lrgs_before_phi_split) { |
|
2016
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1332 insert--; |
|
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1333 } |
| 10111 | 1334 def = split_Rematerialize(def, pred, insert, maxlrg, splits, slidx, lrg2reach, Reachblock, false); |
| 1335 if (!def) { | |
| 1336 return 0; // Bail out | |
| 1337 } | |
| 0 | 1338 } |
| 1339 // Update the Phi's input edge array | |
| 1340 phi->set_req(i,def); | |
| 1341 // Grab the UP/DOWN sense for the input | |
| 1342 u1 = UP[pidx][slidx]; | |
| 1343 if( u1 != (phi_up != 0)) { | |
| 1344 maxlrg = split_USE(def, b, phi, i, maxlrg, !u1, false, splits,slidx); | |
| 1345 // If it wasn't split bail | |
| 1346 if (!maxlrg) { | |
| 1347 return 0; | |
| 1348 } | |
| 1349 } | |
| 1350 } // End for all inputs to the Phi | |
| 1351 } // End for all Phi Nodes | |
| 1352 // Update _maxlrg to save Union asserts | |
| 10111 | 1353 _lrg_map.set_max_lrg_id(maxlrg); |
| 0 | 1354 |
| 1355 | |
| 1356 //----------PASS 3---------- | |
| 1357 // Pass over all Phi's to union the live ranges | |
| 1358 for( insidx = 0; insidx < phis->size(); insidx++ ) { | |
| 1359 Node *phi = phis->at(insidx); | |
| 1360 assert(phi->is_Phi(),"This list must only contain Phi Nodes"); | |
| 1361 // Walk all inputs to Phi and Union input live range with Phi live range | |
| 1362 for( uint i = 1; i < phi->req(); i++ ) { | |
| 1363 // Grab the input node | |
| 1364 Node *n = phi->in(i); | |
| 10111 | 1365 assert(n, "node should exist"); |
| 1366 uint lidx = _lrg_map.find(n); | |
| 1367 uint pidx = _lrg_map.find(phi); | |
| 1368 if (lidx < pidx) { | |
| 0 | 1369 Union(n, phi); |
| 10111 | 1370 } |
| 1371 else if(lidx > pidx) { | |
| 0 | 1372 Union(phi, n); |
| 10111 | 1373 } |
| 0 | 1374 } // End for all inputs to the Phi Node |
| 1375 } // End for all Phi Nodes | |
| 1376 // Now union all two address instructions | |
| 10111 | 1377 for (insidx = 0; insidx < defs->size(); insidx++) { |
| 0 | 1378 // Grab the def |
| 1379 n1 = defs->at(insidx); | |
| 1380 // Set new lidx for DEF & handle 2-addr instructions | |
| 10111 | 1381 if (n1->is_Mach() && ((twoidx = n1->as_Mach()->two_adr()) != 0)) { |
| 1382 assert(_lrg_map.find(n1->in(twoidx)) < maxlrg,"Assigning bad live range index"); | |
| 0 | 1383 // Union the input and output live ranges |
| 10111 | 1384 uint lr1 = _lrg_map.find(n1); |
| 1385 uint lr2 = _lrg_map.find(n1->in(twoidx)); | |
| 1386 if (lr1 < lr2) { | |
| 0 | 1387 Union(n1, n1->in(twoidx)); |
| 10111 | 1388 } |
| 1389 else if (lr1 > lr2) { | |
| 0 | 1390 Union(n1->in(twoidx), n1); |
| 10111 | 1391 } |
| 0 | 1392 } // End if two address |
| 1393 } // End for all defs | |
| 1394 // DEBUG | |
| 1395 #ifdef ASSERT | |
| 1396 // Validate all live range index assignments | |
| 10111 | 1397 for (bidx = 0; bidx < _cfg._num_blocks; bidx++) { |
| 0 | 1398 b = _cfg._blocks[bidx]; |
| 10111 | 1399 for (insidx = 0; insidx <= b->end_idx(); insidx++) { |
| 0 | 1400 Node *n = b->_nodes[insidx]; |
| 10111 | 1401 uint defidx = _lrg_map.find(n); |
| 1402 assert(defidx < _lrg_map.max_lrg_id(), "Bad live range index in Split"); | |
| 0 | 1403 assert(defidx < maxlrg,"Bad live range index in Split"); |
| 1404 } | |
| 1405 } | |
| 1406 // Issue a warning if splitting made no progress | |
| 1407 int noprogress = 0; | |
| 10111 | 1408 for (slidx = 0; slidx < spill_cnt; slidx++) { |
| 1409 if (PrintOpto && WizardMode && splits.at(slidx) == 0) { | |
| 0 | 1410 tty->print_cr("Failed to split live range %d", lidxs.at(slidx)); |
| 1411 //BREAKPOINT; | |
| 1412 } | |
| 1413 else { | |
| 1414 noprogress++; | |
| 1415 } | |
| 1416 } | |
| 1417 if(!noprogress) { | |
| 1418 tty->print_cr("Failed to make progress in Split"); | |
| 1419 //BREAKPOINT; | |
| 1420 } | |
| 1421 #endif | |
| 1422 // Return updated count of live ranges | |
| 1423 return maxlrg; | |
| 1424 } |