Mon, 25 May 2020 14:24:27 +0800
8244407: JVM crashes after transformation in C2 IdealLoopTree::split_fall_in
Reviewed-by: thartmann, kvn, andrew
Contributed-by: zhouyong44@huawei.com
1 /*
2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
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 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #ifndef SHARE_VM_OPTO_REGMASK_HPP
26 #define SHARE_VM_OPTO_REGMASK_HPP
28 #include "code/vmreg.hpp"
29 #include "libadt/port.hpp"
30 #include "opto/optoreg.hpp"
31 #if defined ADGLOBALS_MD_HPP
32 # include ADGLOBALS_MD_HPP
33 #elif defined TARGET_ARCH_MODEL_x86_32
34 # include "adfiles/adGlobals_x86_32.hpp"
35 #elif defined TARGET_ARCH_MODEL_x86_64
36 # include "adfiles/adGlobals_x86_64.hpp"
37 #elif defined TARGET_ARCH_MODEL_sparc
38 # include "adfiles/adGlobals_sparc.hpp"
39 #elif defined TARGET_ARCH_MODEL_zero
40 # include "adfiles/adGlobals_zero.hpp"
41 #elif defined TARGET_ARCH_MODEL_ppc_64
42 # include "adfiles/adGlobals_ppc_64.hpp"
43 #endif
45 // Some fun naming (textual) substitutions:
46 //
47 // RegMask::get_low_elem() ==> RegMask::find_first_elem()
48 // RegMask::Special ==> RegMask::Empty
49 // RegMask::_flags ==> RegMask::is_AllStack()
50 // RegMask::operator<<=() ==> RegMask::Insert()
51 // RegMask::operator>>=() ==> RegMask::Remove()
52 // RegMask::Union() ==> RegMask::OR
53 // RegMask::Inter() ==> RegMask::AND
54 //
55 // OptoRegister::RegName ==> OptoReg::Name
56 //
57 // OptoReg::stack0() ==> _last_Mach_Reg or ZERO in core version
58 //
59 // numregs in chaitin ==> proper degree in chaitin
61 //-------------Non-zero bit search methods used by RegMask---------------------
62 // Find lowest 1, or return 32 if empty
63 int find_lowest_bit( uint32 mask );
64 // Find highest 1, or return 32 if empty
65 int find_hihghest_bit( uint32 mask );
67 //------------------------------RegMask----------------------------------------
68 // The ADL file describes how to print the machine-specific registers, as well
69 // as any notion of register classes. We provide a register mask, which is
70 // just a collection of Register numbers.
72 // The ADLC defines 2 macros, RM_SIZE and FORALL_BODY.
73 // RM_SIZE is the size of a register mask in words.
74 // FORALL_BODY replicates a BODY macro once per word in the register mask.
75 // The usage is somewhat clumsy and limited to the regmask.[h,c]pp files.
76 // However, it means the ADLC can redefine the unroll macro and all loops
77 // over register masks will be unrolled by the correct amount.
79 class RegMask VALUE_OBJ_CLASS_SPEC {
80 union {
81 double _dummy_force_double_alignment[RM_SIZE>>1];
82 // Array of Register Mask bits. This array is large enough to cover
83 // all the machine registers and all parameters that need to be passed
84 // on the stack (stack registers) up to some interesting limit. Methods
85 // that need more parameters will NOT be compiled. On Intel, the limit
86 // is something like 90+ parameters.
87 int _A[RM_SIZE];
88 };
90 enum {
91 _WordBits = BitsPerInt,
92 _LogWordBits = LogBitsPerInt,
93 _RM_SIZE = RM_SIZE // local constant, imported, then hidden by #undef
94 };
96 public:
97 enum { CHUNK_SIZE = RM_SIZE*_WordBits };
99 // SlotsPerLong is 2, since slots are 32 bits and longs are 64 bits.
100 // Also, consider the maximum alignment size for a normally allocated
101 // value. Since we allocate register pairs but not register quads (at
102 // present), this alignment is SlotsPerLong (== 2). A normally
103 // aligned allocated register is either a single register, or a pair
104 // of adjacent registers, the lower-numbered being even.
105 // See also is_aligned_Pairs() below, and the padding added before
106 // Matcher::_new_SP to keep allocated pairs aligned properly.
107 // If we ever go to quad-word allocations, SlotsPerQuad will become
108 // the controlling alignment constraint. Note that this alignment
109 // requirement is internal to the allocator, and independent of any
110 // particular platform.
111 enum { SlotsPerLong = 2,
112 SlotsPerVecS = 1,
113 SlotsPerVecD = 2,
114 SlotsPerVecX = 4,
115 SlotsPerVecY = 8 };
117 // A constructor only used by the ADLC output. All mask fields are filled
118 // in directly. Calls to this look something like RM(1,2,3,4);
119 RegMask(
120 # define BODY(I) int a##I,
121 FORALL_BODY
122 # undef BODY
123 int dummy = 0 ) {
124 # define BODY(I) _A[I] = a##I;
125 FORALL_BODY
126 # undef BODY
127 }
129 // Handy copying constructor
130 RegMask( RegMask *rm ) {
131 # define BODY(I) _A[I] = rm->_A[I];
132 FORALL_BODY
133 # undef BODY
134 }
136 // Construct an empty mask
137 RegMask( ) { Clear(); }
139 // Construct a mask with a single bit
140 RegMask( OptoReg::Name reg ) { Clear(); Insert(reg); }
142 // Check for register being in mask
143 int Member( OptoReg::Name reg ) const {
144 assert( reg < CHUNK_SIZE, "" );
145 return _A[reg>>_LogWordBits] & (1<<(reg&(_WordBits-1)));
146 }
148 // The last bit in the register mask indicates that the mask should repeat
149 // indefinitely with ONE bits. Returns TRUE if mask is infinite or
150 // unbounded in size. Returns FALSE if mask is finite size.
151 int is_AllStack() const { return _A[RM_SIZE-1] >> (_WordBits-1); }
153 // Work around an -xO3 optimization problme in WS6U1. The old way:
154 // void set_AllStack() { _A[RM_SIZE-1] |= (1<<(_WordBits-1)); }
155 // will cause _A[RM_SIZE-1] to be clobbered, not updated when set_AllStack()
156 // follows an Insert() loop, like the one found in init_spill_mask(). Using
157 // Insert() instead works because the index into _A in computed instead of
158 // constant. See bug 4665841.
159 void set_AllStack() { Insert(OptoReg::Name(CHUNK_SIZE-1)); }
161 // Test for being a not-empty mask.
162 int is_NotEmpty( ) const {
163 int tmp = 0;
164 # define BODY(I) tmp |= _A[I];
165 FORALL_BODY
166 # undef BODY
167 return tmp;
168 }
170 // Find lowest-numbered register from mask, or BAD if mask is empty.
171 OptoReg::Name find_first_elem() const {
172 int base, bits;
173 # define BODY(I) if( (bits = _A[I]) != 0 ) base = I<<_LogWordBits; else
174 FORALL_BODY
175 # undef BODY
176 { base = OptoReg::Bad; bits = 1<<0; }
177 return OptoReg::Name(base + find_lowest_bit(bits));
178 }
179 // Get highest-numbered register from mask, or BAD if mask is empty.
180 OptoReg::Name find_last_elem() const {
181 int base, bits;
182 # define BODY(I) if( (bits = _A[RM_SIZE-1-I]) != 0 ) base = (RM_SIZE-1-I)<<_LogWordBits; else
183 FORALL_BODY
184 # undef BODY
185 { base = OptoReg::Bad; bits = 1<<0; }
186 return OptoReg::Name(base + find_hihghest_bit(bits));
187 }
189 // Find the lowest-numbered register pair in the mask. Return the
190 // HIGHEST register number in the pair, or BAD if no pairs.
191 // Assert that the mask contains only bit pairs.
192 OptoReg::Name find_first_pair() const;
194 // Clear out partial bits; leave only aligned adjacent bit pairs.
195 void clear_to_pairs();
196 // Smear out partial bits; leave only aligned adjacent bit pairs.
197 void smear_to_pairs();
198 // Verify that the mask contains only aligned adjacent bit pairs
199 void verify_pairs() const { assert( is_aligned_pairs(), "mask is not aligned, adjacent pairs" ); }
200 // Test that the mask contains only aligned adjacent bit pairs
201 bool is_aligned_pairs() const;
203 // mask is a pair of misaligned registers
204 bool is_misaligned_pair() const { return Size()==2 && !is_aligned_pairs(); }
205 // Test for single register
206 int is_bound1() const;
207 // Test for a single adjacent pair
208 int is_bound_pair() const;
209 // Test for a single adjacent set of ideal register's size.
210 int is_bound(uint ireg) const {
211 if (is_vector(ireg)) {
212 if (is_bound_set(num_registers(ireg)))
213 return true;
214 } else if (is_bound1() || is_bound_pair()) {
215 return true;
216 }
217 return false;
218 }
220 // Find the lowest-numbered register set in the mask. Return the
221 // HIGHEST register number in the set, or BAD if no sets.
222 // Assert that the mask contains only bit sets.
223 OptoReg::Name find_first_set(const int size) const;
225 // Clear out partial bits; leave only aligned adjacent bit sets of size.
226 void clear_to_sets(const int size);
227 // Smear out partial bits to aligned adjacent bit sets.
228 void smear_to_sets(const int size);
229 // Verify that the mask contains only aligned adjacent bit sets
230 void verify_sets(int size) const { assert(is_aligned_sets(size), "mask is not aligned, adjacent sets"); }
231 // Test that the mask contains only aligned adjacent bit sets
232 bool is_aligned_sets(const int size) const;
234 // mask is a set of misaligned registers
235 bool is_misaligned_set(int size) const { return (int)Size()==size && !is_aligned_sets(size);}
237 // Test for a single adjacent set
238 int is_bound_set(const int size) const;
240 static bool is_vector(uint ireg);
241 static int num_registers(uint ireg);
243 // Fast overlap test. Non-zero if any registers in common.
244 int overlap( const RegMask &rm ) const {
245 return
246 # define BODY(I) (_A[I] & rm._A[I]) |
247 FORALL_BODY
248 # undef BODY
249 0 ;
250 }
252 // Special test for register pressure based splitting
253 // UP means register only, Register plus stack, or stack only is DOWN
254 bool is_UP() const;
256 // Clear a register mask
257 void Clear( ) {
258 # define BODY(I) _A[I] = 0;
259 FORALL_BODY
260 # undef BODY
261 }
263 // Fill a register mask with 1's
264 void Set_All( ) {
265 # define BODY(I) _A[I] = -1;
266 FORALL_BODY
267 # undef BODY
268 }
270 // Insert register into mask
271 void Insert( OptoReg::Name reg ) {
272 assert( reg < CHUNK_SIZE, "" );
273 _A[reg>>_LogWordBits] |= (1<<(reg&(_WordBits-1)));
274 }
276 // Remove register from mask
277 void Remove( OptoReg::Name reg ) {
278 assert( reg < CHUNK_SIZE, "" );
279 _A[reg>>_LogWordBits] &= ~(1<<(reg&(_WordBits-1)));
280 }
282 // OR 'rm' into 'this'
283 void OR( const RegMask &rm ) {
284 # define BODY(I) this->_A[I] |= rm._A[I];
285 FORALL_BODY
286 # undef BODY
287 }
289 // AND 'rm' into 'this'
290 void AND( const RegMask &rm ) {
291 # define BODY(I) this->_A[I] &= rm._A[I];
292 FORALL_BODY
293 # undef BODY
294 }
296 // Subtract 'rm' from 'this'
297 void SUBTRACT( const RegMask &rm ) {
298 # define BODY(I) _A[I] &= ~rm._A[I];
299 FORALL_BODY
300 # undef BODY
301 }
303 // Compute size of register mask: number of bits
304 uint Size() const;
306 #ifndef PRODUCT
307 void print() const { dump(); }
308 void dump(outputStream *st = tty) const; // Print a mask
309 #endif
311 static const RegMask Empty; // Common empty mask
313 static bool can_represent(OptoReg::Name reg) {
314 // NOTE: -1 in computation reflects the usage of the last
315 // bit of the regmask as an infinite stack flag and
316 // -7 is to keep mask aligned for largest value (VecY).
317 return (int)reg < (int)(CHUNK_SIZE-1);
318 }
319 static bool can_represent_arg(OptoReg::Name reg) {
320 // NOTE: -SlotsPerVecY in computation reflects the need
321 // to keep mask aligned for largest value (VecY).
322 return (int)reg < (int)(CHUNK_SIZE-SlotsPerVecY);
323 }
324 };
326 // Do not use this constant directly in client code!
327 #undef RM_SIZE
329 #endif // SHARE_VM_OPTO_REGMASK_HPP