1 /*
2 * Copyright (c) 1997, 2019, 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 */
24
25 #ifndef SHARE_OPTO_REGMASK_HPP
26 #define SHARE_OPTO_REGMASK_HPP
27
28 #include "code/vmreg.hpp"
29 #include "opto/optoreg.hpp"
30 #include "utilities/count_leading_zeros.hpp"
31 #include "utilities/count_trailing_zeros.hpp"
32
33 // Some fun naming (textual) substitutions:
34 //
35 // RegMask::get_low_elem() ==> RegMask::find_first_elem()
36 // RegMask::Special ==> RegMask::Empty
37 // RegMask::_flags ==> RegMask::is_AllStack()
38 // RegMask::operator<<=() ==> RegMask::Insert()
39 // RegMask::operator>>=() ==> RegMask::Remove()
40 // RegMask::Union() ==> RegMask::OR
41 // RegMask::Inter() ==> RegMask::AND
42 //
43 // OptoRegister::RegName ==> OptoReg::Name
44 //
45 // OptoReg::stack0() ==> _last_Mach_Reg or ZERO in core version
46 //
47 // numregs in chaitin ==> proper degree in chaitin
48
49 //-------------Non-zero bit search methods used by RegMask---------------------
50 // Find lowest 1, undefined if empty/0
51 static int find_lowest_bit(uint32_t mask) {
52 return count_trailing_zeros(mask);
53 }
54 // Find highest 1, undefined if empty/0
55 static int find_highest_bit(uint32_t mask) {
56 return count_leading_zeros(mask) ^ 31;
57 }
58
59 //------------------------------RegMask----------------------------------------
60 // The ADL file describes how to print the machine-specific registers, as well
61 // as any notion of register classes. We provide a register mask, which is
62 // just a collection of Register numbers.
63
64 // The ADLC defines 2 macros, RM_SIZE and FORALL_BODY.
65 // RM_SIZE is the size of a register mask in words.
66 // FORALL_BODY replicates a BODY macro once per word in the register mask.
67 // The usage is somewhat clumsy and limited to the regmask.[h,c]pp files.
68 // However, it means the ADLC can redefine the unroll macro and all loops
69 // over register masks will be unrolled by the correct amount.
70
71 class RegMask {
72 union {
73 double _dummy_force_double_alignment[RM_SIZE>>1];
74 // Array of Register Mask bits. This array is large enough to cover
75 // all the machine registers and all parameters that need to be passed
76 // on the stack (stack registers) up to some interesting limit. Methods
77 // that need more parameters will NOT be compiled. On Intel, the limit
78 // is something like 90+ parameters.
79 int _A[RM_SIZE];
80 };
81
82 enum {
83 _WordBits = BitsPerInt,
84 _LogWordBits = LogBitsPerInt,
85 _RM_SIZE = RM_SIZE // local constant, imported, then hidden by #undef
86 };
87
88 public:
89 enum { CHUNK_SIZE = RM_SIZE*_WordBits };
90
91 // SlotsPerLong is 2, since slots are 32 bits and longs are 64 bits.
92 // Also, consider the maximum alignment size for a normally allocated
93 // value. Since we allocate register pairs but not register quads (at
94 // present), this alignment is SlotsPerLong (== 2). A normally
95 // aligned allocated register is either a single register, or a pair
96 // of adjacent registers, the lower-numbered being even.
97 // See also is_aligned_Pairs() below, and the padding added before
98 // Matcher::_new_SP to keep allocated pairs aligned properly.
99 // If we ever go to quad-word allocations, SlotsPerQuad will become
100 // the controlling alignment constraint. Note that this alignment
101 // requirement is internal to the allocator, and independent of any
102 // particular platform.
103 enum { SlotsPerLong = 2,
104 SlotsPerVecS = 1,
105 SlotsPerVecD = 2,
106 SlotsPerVecX = 4,
107 SlotsPerVecY = 8,
108 SlotsPerVecZ = 16 };
109
110 // A constructor only used by the ADLC output. All mask fields are filled
111 // in directly. Calls to this look something like RM(1,2,3,4);
112 RegMask(
113 # define BODY(I) int a##I,
114 FORALL_BODY
115 # undef BODY
116 int dummy = 0 ) {
117 # define BODY(I) _A[I] = a##I;
118 FORALL_BODY
119 # undef BODY
120 }
121
122 // Handy copying constructor
123 RegMask( RegMask *rm ) {
124 # define BODY(I) _A[I] = rm->_A[I];
125 FORALL_BODY
126 # undef BODY
127 }
128
129 // Construct an empty mask
130 RegMask( ) { Clear(); }
131
132 // Construct a mask with a single bit
133 RegMask( OptoReg::Name reg ) { Clear(); Insert(reg); }
134
135 // Check for register being in mask
136 int Member( OptoReg::Name reg ) const {
137 assert( reg < CHUNK_SIZE, "" );
138 return _A[reg>>_LogWordBits] & (1<<(reg&(_WordBits-1)));
139 }
140
141 // The last bit in the register mask indicates that the mask should repeat
142 // indefinitely with ONE bits. Returns TRUE if mask is infinite or
143 // unbounded in size. Returns FALSE if mask is finite size.
144 int is_AllStack() const { return _A[RM_SIZE-1] >> (_WordBits-1); }
145
146 // Work around an -xO3 optimization problme in WS6U1. The old way:
147 // void set_AllStack() { _A[RM_SIZE-1] |= (1<<(_WordBits-1)); }
148 // will cause _A[RM_SIZE-1] to be clobbered, not updated when set_AllStack()
149 // follows an Insert() loop, like the one found in init_spill_mask(). Using
150 // Insert() instead works because the index into _A in computed instead of
151 // constant. See bug 4665841.
152 void set_AllStack() { Insert(OptoReg::Name(CHUNK_SIZE-1)); }
153
154 // Test for being a not-empty mask.
155 int is_NotEmpty( ) const {
156 int tmp = 0;
157 # define BODY(I) tmp |= _A[I];
158 FORALL_BODY
159 # undef BODY
160 return tmp;
161 }
162
163 // Find lowest-numbered register from mask, or BAD if mask is empty.
164 OptoReg::Name find_first_elem() const {
165 int base, bits;
166 # define BODY(I) if( (bits = _A[I]) != 0 ) base = I<<_LogWordBits; else
167 FORALL_BODY
168 # undef BODY
169 { base = OptoReg::Bad; bits = 1<<0; }
170 return OptoReg::Name(base + find_lowest_bit(bits));
171 }
172 // Get highest-numbered register from mask, or BAD if mask is empty.
173 OptoReg::Name find_last_elem() const {
174 int base, bits;
175 # define BODY(I) if( (bits = _A[RM_SIZE-1-I]) != 0 ) base = (RM_SIZE-1-I)<<_LogWordBits; else
176 FORALL_BODY
177 # undef BODY
178 { base = OptoReg::Bad; bits = 1<<0; }
179 return OptoReg::Name(base + find_highest_bit(bits));
180 }
181
182 // Clear out partial bits; leave only aligned adjacent bit pairs.
183 void clear_to_pairs();
184 // Verify that the mask contains only aligned adjacent bit pairs
185 void verify_pairs() const { assert( is_aligned_pairs(), "mask is not aligned, adjacent pairs" ); }
186 // Test that the mask contains only aligned adjacent bit pairs
187 bool is_aligned_pairs() const;
188
189 // mask is a pair of misaligned registers
190 bool is_misaligned_pair() const { return Size()==2 && !is_aligned_pairs(); }
191 // Test for single register
192 int is_bound1() const;
193 // Test for a single adjacent pair
194 int is_bound_pair() const;
195 // Test for a single adjacent set of ideal register's size.
196 int is_bound(uint ireg) const {
197 if (is_vector(ireg)) {
198 if (is_bound_set(num_registers(ireg)))
199 return true;
200 } else if (is_bound1() || is_bound_pair()) {
201 return true;
202 }
203 return false;
204 }
205
206 // Find the lowest-numbered register set in the mask. Return the
207 // HIGHEST register number in the set, or BAD if no sets.
208 // Assert that the mask contains only bit sets.
209 OptoReg::Name find_first_set(const int size) const;
210
211 // Clear out partial bits; leave only aligned adjacent bit sets of size.
212 void clear_to_sets(const int size);
213 // Smear out partial bits to aligned adjacent bit sets.
214 void smear_to_sets(const int size);
215 // Verify that the mask contains only aligned adjacent bit sets
216 void verify_sets(int size) const { assert(is_aligned_sets(size), "mask is not aligned, adjacent sets"); }
217 // Test that the mask contains only aligned adjacent bit sets
218 bool is_aligned_sets(const int size) const;
219
220 // Test for a single adjacent set
221 int is_bound_set(const int size) const;
222
223 static bool is_vector(uint ireg);
224 static int num_registers(uint ireg);
225
226 // Fast overlap test. Non-zero if any registers in common.
227 int overlap( const RegMask &rm ) const {
228 return
229 # define BODY(I) (_A[I] & rm._A[I]) |
230 FORALL_BODY
231 # undef BODY
232 0 ;
233 }
234
235 // Special test for register pressure based splitting
236 // UP means register only, Register plus stack, or stack only is DOWN
237 bool is_UP() const;
238
239 // Clear a register mask
240 void Clear( ) {
241 # define BODY(I) _A[I] = 0;
242 FORALL_BODY
243 # undef BODY
244 }
245
246 // Fill a register mask with 1's
247 void Set_All( ) {
248 # define BODY(I) _A[I] = -1;
249 FORALL_BODY
250 # undef BODY
251 }
252
253 // Insert register into mask
254 void Insert( OptoReg::Name reg ) {
255 assert( reg < CHUNK_SIZE, "" );
256 _A[reg>>_LogWordBits] |= (1<<(reg&(_WordBits-1)));
257 }
258
259 // Remove register from mask
260 void Remove( OptoReg::Name reg ) {
261 assert( reg < CHUNK_SIZE, "" );
262 _A[reg>>_LogWordBits] &= ~(1<<(reg&(_WordBits-1)));
263 }
264
265 // OR 'rm' into 'this'
266 void OR( const RegMask &rm ) {
267 # define BODY(I) this->_A[I] |= rm._A[I];
268 FORALL_BODY
269 # undef BODY
270 }
271
272 // AND 'rm' into 'this'
273 void AND( const RegMask &rm ) {
274 # define BODY(I) this->_A[I] &= rm._A[I];
275 FORALL_BODY
276 # undef BODY
277 }
278
279 // Subtract 'rm' from 'this'
280 void SUBTRACT( const RegMask &rm ) {
281 # define BODY(I) _A[I] &= ~rm._A[I];
282 FORALL_BODY
283 # undef BODY
284 }
285
286 // Compute size of register mask: number of bits
287 uint Size() const;
288
289 #ifndef PRODUCT
290 void print() const { dump(); }
291 void dump(outputStream *st = tty) const; // Print a mask
292 #endif
293
294 static const RegMask Empty; // Common empty mask
295
296 static bool can_represent(OptoReg::Name reg) {
297 // NOTE: -1 in computation reflects the usage of the last
298 // bit of the regmask as an infinite stack flag and
299 // -7 is to keep mask aligned for largest value (VecZ).
300 return (int)reg < (int)(CHUNK_SIZE-1);
301 }
302 static bool can_represent_arg(OptoReg::Name reg) {
303 // NOTE: -SlotsPerVecZ in computation reflects the need
304 // to keep mask aligned for largest value (VecZ).
305 return (int)reg < (int)(CHUNK_SIZE-SlotsPerVecZ);
306 }
307 };
308
309 // Do not use this constant directly in client code!
310 #undef RM_SIZE
311
312 #endif // SHARE_OPTO_REGMASK_HPP