1 /*
  2  * Copyright (c) 1997, 2020, 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 #include "precompiled.hpp"
 26 #include "opto/ad.hpp"
 27 #include "opto/chaitin.hpp"
 28 #include "opto/compile.hpp"
 29 #include "opto/matcher.hpp"
 30 #include "opto/node.hpp"
 31 #include "opto/regmask.hpp"
 32 #include "utilities/population_count.hpp"
 33 #include "utilities/powerOfTwo.hpp"
 34 
 35 #define RM_SIZE _RM_SIZE /* a constant private to the class RegMask */
 36 
 37 //------------------------------dump-------------------------------------------
 38 
 39 #ifndef PRODUCT
 40 void OptoReg::dump(int r, outputStream *st) {
 41   switch (r) {
 42   case Special: st->print("r---"); break;
 43   case Bad:     st->print("rBAD"); break;
 44   default:
 45     if (r < _last_Mach_Reg) st->print("%s", Matcher::regName[r]);
 46     else st->print("rS%d",r);
 47     break;
 48   }
 49 }
 50 #endif
 51 
 52 
 53 //=============================================================================
 54 const RegMask RegMask::Empty(
 55 # define BODY(I) 0,
 56   FORALL_BODY
 57 # undef BODY
 58   0
 59 );
 60 
 61 //=============================================================================
 62 bool RegMask::is_vector(uint ireg) {
 63   return (ireg == Op_VecA || ireg == Op_VecS || ireg == Op_VecD ||
 64           ireg == Op_VecX || ireg == Op_VecY || ireg == Op_VecZ );
 65 }
 66 
 67 int RegMask::num_registers(uint ireg) {
 68     switch(ireg) {
 69       case Op_VecZ:
 70         return SlotsPerVecZ;
 71       case Op_VecY:
 72         return SlotsPerVecY;
 73       case Op_VecX:
 74         return SlotsPerVecX;
 75       case Op_VecD:
 76         return SlotsPerVecD;
 77       case Op_RegD:
 78       case Op_RegL:
 79 #ifdef _LP64
 80       case Op_RegP:
 81 #endif
 82         return 2;
 83       case Op_VecA:
 84         assert(Matcher::supports_scalable_vector(), "does not support scalable vector");
 85         return SlotsPerVecA;
 86     }
 87     // Op_VecS and the rest ideal registers.
 88     return 1;
 89 }
 90 
 91 int RegMask::num_registers(uint ireg, LRG &lrg) {
 92   int n_regs = num_registers(ireg);
 93 
 94   // assigned is OptoReg which is selected by register allocator
 95   OptoReg::Name assigned = lrg.reg();
 96   assert(OptoReg::is_valid(assigned), "should be valid opto register");
 97 
 98   if (lrg.is_scalable() && OptoReg::is_stack(assigned)) {
 99     n_regs = lrg.scalable_reg_slots();
100   }
101   return n_regs;
102 }
103 
104 // Clear out partial bits; leave only bit pairs
105 void RegMask::clear_to_pairs() {
106   assert(valid_watermarks(), "sanity");
107   for (int i = _lwm; i <= _hwm; i++) {
108     int bits = _A[i];
109     bits &= ((bits & 0x55555555)<<1); // 1 hi-bit set for each pair
110     bits |= (bits>>1);          // Smear 1 hi-bit into a pair
111     _A[i] = bits;
112   }
113   assert(is_aligned_pairs(), "mask is not aligned, adjacent pairs");
114 }
115 
116 bool RegMask::is_misaligned_pair() const {
117   return Size() == 2 && !is_aligned_pairs();
118 }
119 
120 bool RegMask::is_aligned_pairs() const {
121   // Assert that the register mask contains only bit pairs.
122   assert(valid_watermarks(), "sanity");
123   for (int i = _lwm; i <= _hwm; i++) {
124     int bits = _A[i];
125     while (bits) {              // Check bits for pairing
126       int bit = bits & -bits;   // Extract low bit
127       // Low bit is not odd means its mis-aligned.
128       if ((bit & 0x55555555) == 0) return false;
129       bits -= bit;              // Remove bit from mask
130       // Check for aligned adjacent bit
131       if ((bits & (bit<<1)) == 0) return false;
132       bits -= (bit<<1);         // Remove other halve of pair
133     }
134   }
135   return true;
136 }
137 
138 // Return TRUE if the mask contains a single bit
139 bool RegMask::is_bound1() const {
140   if (is_AllStack()) return false;
141   return Size() == 1;
142 }
143 
144 // Return TRUE if the mask contains an adjacent pair of bits and no other bits.
145 bool RegMask::is_bound_pair() const {
146   if (is_AllStack()) return false;
147   int bit = -1;                 // Set to hold the one bit allowed
148   assert(valid_watermarks(), "sanity");
149   for (int i = _lwm; i <= _hwm; i++) {
150     if (_A[i]) {                   // Found some bits
151       if (bit != -1) return false; // Already had bits, so fail
152       bit = _A[i] & -(_A[i]);      // Extract 1 bit from mask
153       if ((bit << 1) != 0) {       // Bit pair stays in same word?
154         if ((bit | (bit<<1)) != _A[i])
155           return false;            // Require adjacent bit pair and no more bits
156       } else {                     // Else its a split-pair case
157         if(bit != _A[i]) return false; // Found many bits, so fail
158         i++;                       // Skip iteration forward
159         if (i > _hwm || _A[i] != 1)
160           return false; // Require 1 lo bit in next word
161       }
162     }
163   }
164   // True for both the empty mask and for a bit pair
165   return true;
166 }
167 
168 // Test for a single adjacent set of ideal register's size.
169 bool RegMask::is_bound(uint ireg) const {
170   if (is_vector(ireg)) {
171     if (is_bound_set(num_registers(ireg)))
172       return true;
173   } else if (is_bound1() || is_bound_pair()) {
174     return true;
175   }
176   return false;
177 }
178 // Check that whether given reg number with size is valid
179 // for current regmask, where reg is the highest number.
180 bool RegMask::is_valid_reg(OptoReg::Name reg, const int size) const {
181   for (int i = 0; i < size; i++) {
182     if (!Member(reg - i)) {
183       return false;
184     }
185   }
186   return true;
187 }
188 
189 // only indicies of power 2 are accessed, so index 3 is only filled in for storage.
190 static int low_bits[5] = { 0x55555555, 0x11111111, 0x01010101, 0x00000000, 0x00010001 };
191 
192 // Find the lowest-numbered register set in the mask.  Return the
193 // HIGHEST register number in the set, or BAD if no sets.
194 // Works also for size 1.
195 OptoReg::Name RegMask::find_first_set(LRG &lrg, const int size) const {
196   if (lrg.is_scalable()) {
197     // For scalable vector register, regmask is SlotsPerVecA bits aligned.
198     assert(is_aligned_sets(SlotsPerVecA), "mask is not aligned, adjacent sets");
199   } else {
200     assert(is_aligned_sets(size), "mask is not aligned, adjacent sets");
201   }
202   assert(valid_watermarks(), "sanity");
203   for (int i = _lwm; i <= _hwm; i++) {
204     if (_A[i]) {                // Found some bits
205       // Convert to bit number, return hi bit in pair
206       return OptoReg::Name((i<<_LogWordBits) + find_lowest_bit(_A[i]) + (size - 1));
207     }
208   }
209   return OptoReg::Bad;
210 }
211 
212 // Clear out partial bits; leave only aligned adjacent bit pairs
213 void RegMask::clear_to_sets(const int size) {
214   if (size == 1) return;
215   assert(2 <= size && size <= 16, "update low bits table");
216   assert(is_power_of_2(size), "sanity");
217   assert(valid_watermarks(), "sanity");
218   int low_bits_mask = low_bits[size>>2];
219   for (int i = _lwm; i <= _hwm; i++) {
220     int bits = _A[i];
221     int sets = (bits & low_bits_mask);
222     for (int j = 1; j < size; j++) {
223       sets = (bits & (sets<<1)); // filter bits which produce whole sets
224     }
225     sets |= (sets>>1);           // Smear 1 hi-bit into a set
226     if (size > 2) {
227       sets |= (sets>>2);         // Smear 2 hi-bits into a set
228       if (size > 4) {
229         sets |= (sets>>4);       // Smear 4 hi-bits into a set
230         if (size > 8) {
231           sets |= (sets>>8);     // Smear 8 hi-bits into a set
232         }
233       }
234     }
235     _A[i] = sets;
236   }
237   assert(is_aligned_sets(size), "mask is not aligned, adjacent sets");
238 }
239 
240 // Smear out partial bits to aligned adjacent bit sets
241 void RegMask::smear_to_sets(const int size) {
242   if (size == 1) return;
243   assert(2 <= size && size <= 16, "update low bits table");
244   assert(is_power_of_2(size), "sanity");
245   assert(valid_watermarks(), "sanity");
246   int low_bits_mask = low_bits[size>>2];
247   for (int i = _lwm; i <= _hwm; i++) {
248     int bits = _A[i];
249     int sets = 0;
250     for (int j = 0; j < size; j++) {
251       sets |= (bits & low_bits_mask);  // collect partial bits
252       bits  = bits>>1;
253     }
254     sets |= (sets<<1);           // Smear 1 lo-bit  into a set
255     if (size > 2) {
256       sets |= (sets<<2);         // Smear 2 lo-bits into a set
257       if (size > 4) {
258         sets |= (sets<<4);       // Smear 4 lo-bits into a set
259         if (size > 8) {
260           sets |= (sets<<8);     // Smear 8 lo-bits into a set
261         }
262       }
263     }
264     _A[i] = sets;
265   }
266   assert(is_aligned_sets(size), "mask is not aligned, adjacent sets");
267 }
268 
269 // Assert that the register mask contains only bit sets.
270 bool RegMask::is_aligned_sets(const int size) const {
271   if (size == 1) return true;
272   assert(2 <= size && size <= 16, "update low bits table");
273   assert(is_power_of_2(size), "sanity");
274   int low_bits_mask = low_bits[size>>2];
275   assert(valid_watermarks(), "sanity");
276   for (int i = _lwm; i <= _hwm; i++) {
277     int bits = _A[i];
278     while (bits) {              // Check bits for pairing
279       int bit = bits & -bits;   // Extract low bit
280       // Low bit is not odd means its mis-aligned.
281       if ((bit & low_bits_mask) == 0) {
282         return false;
283       }
284       // Do extra work since (bit << size) may overflow.
285       int hi_bit = bit << (size-1); // high bit
286       int set = hi_bit + ((hi_bit-1) & ~(bit-1));
287       // Check for aligned adjacent bits in this set
288       if ((bits & set) != set) {
289         return false;
290       }
291       bits -= set;  // Remove this set
292     }
293   }
294   return true;
295 }
296 
297 // Return TRUE if the mask contains one adjacent set of bits and no other bits.
298 // Works also for size 1.
299 int RegMask::is_bound_set(const int size) const {
300   if (is_AllStack()) return false;
301   assert(1 <= size && size <= 16, "update low bits table");
302   assert(valid_watermarks(), "sanity");
303   int bit = -1;                 // Set to hold the one bit allowed
304   for (int i = _lwm; i <= _hwm; i++) {
305     if (_A[i] ) {               // Found some bits
306       if (bit != -1)
307        return false;            // Already had bits, so fail
308       bit = _A[i] & -_A[i];     // Extract low bit from mask
309       int hi_bit = bit << (size-1); // high bit
310       if (hi_bit != 0) {        // Bit set stays in same word?
311         int set = hi_bit + ((hi_bit-1) & ~(bit-1));
312         if (set != _A[i])
313           return false;         // Require adjacent bit set and no more bits
314       } else {                  // Else its a split-set case
315         if (((-1) & ~(bit-1)) != _A[i])
316           return false;         // Found many bits, so fail
317         i++;                    // Skip iteration forward and check high part
318         // The lower (32-size) bits should be 0 since it is split case.
319         int clear_bit_size = 32-size;
320         int shift_back_size = 32-clear_bit_size;
321         int set = bit>>clear_bit_size;
322         set = set & -set; // Remove sign extension.
323         set = (((set << size) - 1) >> shift_back_size);
324         if (i > _hwm || _A[i] != set)
325           return false; // Require expected low bits in next word
326       }
327     }
328   }
329   // True for both the empty mask and for a bit set
330   return true;
331 }
332 
333 // UP means register only, Register plus stack, or stack only is DOWN
334 bool RegMask::is_UP() const {
335   // Quick common case check for DOWN (any stack slot is legal)
336   if (is_AllStack())
337     return false;
338   // Slower check for any stack bits set (also DOWN)
339   if (overlap(Matcher::STACK_ONLY_mask))
340     return false;
341   // Not DOWN, so must be UP
342   return true;
343 }
344 
345 // Compute size of register mask in bits
346 uint RegMask::Size() const {
347   uint sum = 0;
348   assert(valid_watermarks(), "sanity");
349   for (int i = _lwm; i <= _hwm; i++) {
350     sum += population_count((unsigned)_A[i]);
351   }
352   return sum;
353 }
354 
355 #ifndef PRODUCT
356 void RegMask::dump(outputStream *st) const {
357   st->print("[");
358   RegMask rm = *this;           // Structure copy into local temp
359 
360   OptoReg::Name start = rm.find_first_elem(); // Get a register
361   if (OptoReg::is_valid(start)) { // Check for empty mask
362     rm.Remove(start);           // Yank from mask
363     OptoReg::dump(start, st);   // Print register
364     OptoReg::Name last = start;
365 
366     // Now I have printed an initial register.
367     // Print adjacent registers as "rX-rZ" instead of "rX,rY,rZ".
368     // Begin looping over the remaining registers.
369     while (1) {                 //
370       OptoReg::Name reg = rm.find_first_elem(); // Get a register
371       if (!OptoReg::is_valid(reg))
372         break;                  // Empty mask, end loop
373       rm.Remove(reg);           // Yank from mask
374 
375       if (last+1 == reg) {      // See if they are adjacent
376         // Adjacent registers just collect into long runs, no printing.
377         last = reg;
378       } else {                  // Ending some kind of run
379         if (start == last) {    // 1-register run; no special printing
380         } else if (start+1 == last) {
381           st->print(",");       // 2-register run; print as "rX,rY"
382           OptoReg::dump(last, st);
383         } else {                // Multi-register run; print as "rX-rZ"
384           st->print("-");
385           OptoReg::dump(last, st);
386         }
387         st->print(",");         // Seperate start of new run
388         start = last = reg;     // Start a new register run
389         OptoReg::dump(start, st); // Print register
390       } // End of if ending a register run or not
391     } // End of while regmask not empty
392 
393     if (start == last) {        // 1-register run; no special printing
394     } else if (start+1 == last) {
395       st->print(",");           // 2-register run; print as "rX,rY"
396       OptoReg::dump(last, st);
397     } else {                    // Multi-register run; print as "rX-rZ"
398       st->print("-");
399       OptoReg::dump(last, st);
400     }
401     if (rm.is_AllStack()) st->print("...");
402   }
403   st->print("]");
404 }
405 #endif