Mon, 20 Aug 2012 09:07:21 -0700
6340864: Implement vectorization optimizations in hotspot-server
Summary: Added asm encoding and mach nodes for vector arithmetic instructions on x86.
Reviewed-by: roland
1 /*
2 * Copyright (c) 2007, 2012, 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 */
24 #include "precompiled.hpp"
25 #include "memory/allocation.inline.hpp"
26 #include "opto/connode.hpp"
27 #include "opto/vectornode.hpp"
29 //------------------------------VectorNode--------------------------------------
31 // Return the vector operator for the specified scalar operation
32 // and vector length. Also used to check if the code generator
33 // supports the vector operation.
34 int VectorNode::opcode(int sopc, uint vlen, BasicType bt) {
35 switch (sopc) {
36 case Op_AddI:
37 switch (bt) {
38 case T_BOOLEAN:
39 case T_BYTE: return Op_AddVB;
40 case T_CHAR:
41 case T_SHORT: return Op_AddVS;
42 case T_INT: return Op_AddVI;
43 }
44 ShouldNotReachHere();
45 case Op_AddL:
46 assert(bt == T_LONG, "must be");
47 return Op_AddVL;
48 case Op_AddF:
49 assert(bt == T_FLOAT, "must be");
50 return Op_AddVF;
51 case Op_AddD:
52 assert(bt == T_DOUBLE, "must be");
53 return Op_AddVD;
54 case Op_SubI:
55 switch (bt) {
56 case T_BOOLEAN:
57 case T_BYTE: return Op_SubVB;
58 case T_CHAR:
59 case T_SHORT: return Op_SubVS;
60 case T_INT: return Op_SubVI;
61 }
62 ShouldNotReachHere();
63 case Op_SubL:
64 assert(bt == T_LONG, "must be");
65 return Op_SubVL;
66 case Op_SubF:
67 assert(bt == T_FLOAT, "must be");
68 return Op_SubVF;
69 case Op_SubD:
70 assert(bt == T_DOUBLE, "must be");
71 return Op_SubVD;
72 case Op_MulI:
73 switch (bt) {
74 case T_BOOLEAN:
75 case T_BYTE: return 0; // Unimplemented
76 case T_CHAR:
77 case T_SHORT: return Op_MulVS;
78 case T_INT: return Matcher::match_rule_supported(Op_MulVI) ? Op_MulVI : 0; // SSE4_1
79 }
80 ShouldNotReachHere();
81 case Op_MulF:
82 assert(bt == T_FLOAT, "must be");
83 return Op_MulVF;
84 case Op_MulD:
85 assert(bt == T_DOUBLE, "must be");
86 return Op_MulVD;
87 case Op_DivF:
88 assert(bt == T_FLOAT, "must be");
89 return Op_DivVF;
90 case Op_DivD:
91 assert(bt == T_DOUBLE, "must be");
92 return Op_DivVD;
93 case Op_LShiftI:
94 switch (bt) {
95 case T_BOOLEAN:
96 case T_BYTE: return Op_LShiftVB;
97 case T_CHAR:
98 case T_SHORT: return Op_LShiftVS;
99 case T_INT: return Op_LShiftVI;
100 }
101 ShouldNotReachHere();
102 case Op_LShiftL:
103 assert(bt == T_LONG, "must be");
104 return Op_LShiftVL;
105 case Op_RShiftI:
106 switch (bt) {
107 case T_BOOLEAN:
108 case T_BYTE: return Op_RShiftVB;
109 case T_CHAR:
110 case T_SHORT: return Op_RShiftVS;
111 case T_INT: return Op_RShiftVI;
112 }
113 ShouldNotReachHere();
114 case Op_RShiftL:
115 assert(bt == T_LONG, "must be");
116 return Op_RShiftVL;
117 case Op_URShiftI:
118 switch (bt) {
119 case T_BOOLEAN:
120 case T_BYTE: return Op_URShiftVB;
121 case T_CHAR:
122 case T_SHORT: return Op_URShiftVS;
123 case T_INT: return Op_URShiftVI;
124 }
125 ShouldNotReachHere();
126 case Op_URShiftL:
127 assert(bt == T_LONG, "must be");
128 return Op_URShiftVL;
129 case Op_AndI:
130 case Op_AndL:
131 return Op_AndV;
132 case Op_OrI:
133 case Op_OrL:
134 return Op_OrV;
135 case Op_XorI:
136 case Op_XorL:
137 return Op_XorV;
139 case Op_LoadB:
140 case Op_LoadUB:
141 case Op_LoadUS:
142 case Op_LoadS:
143 case Op_LoadI:
144 case Op_LoadL:
145 case Op_LoadF:
146 case Op_LoadD:
147 return Op_LoadVector;
149 case Op_StoreB:
150 case Op_StoreC:
151 case Op_StoreI:
152 case Op_StoreL:
153 case Op_StoreF:
154 case Op_StoreD:
155 return Op_StoreVector;
156 }
157 return 0; // Unimplemented
158 }
160 bool VectorNode::implemented(int opc, uint vlen, BasicType bt) {
161 if (is_java_primitive(bt) &&
162 (vlen > 1) && is_power_of_2(vlen) &&
163 Matcher::vector_size_supported(bt, vlen)) {
164 int vopc = VectorNode::opcode(opc, vlen, bt);
165 return vopc > 0 && Matcher::has_match_rule(vopc);
166 }
167 return false;
168 }
170 bool VectorNode::is_shift(Node* n) {
171 switch (n->Opcode()) {
172 case Op_LShiftI:
173 case Op_LShiftL:
174 case Op_RShiftI:
175 case Op_RShiftL:
176 case Op_URShiftI:
177 case Op_URShiftL:
178 return true;
179 }
180 return false;
181 }
183 // Check if input is loop invarient vector.
184 bool VectorNode::is_invariant_vector(Node* n) {
185 // Only Replicate vector nodes are loop invarient for now.
186 switch (n->Opcode()) {
187 case Op_ReplicateB:
188 case Op_ReplicateS:
189 case Op_ReplicateI:
190 case Op_ReplicateL:
191 case Op_ReplicateF:
192 case Op_ReplicateD:
193 return true;
194 }
195 return false;
196 }
198 // Return the vector version of a scalar operation node.
199 VectorNode* VectorNode::make(Compile* C, int opc, Node* n1, Node* n2, uint vlen, BasicType bt) {
200 const TypeVect* vt = TypeVect::make(bt, vlen);
201 int vopc = VectorNode::opcode(opc, vlen, bt);
203 switch (vopc) {
204 case Op_AddVB: return new (C, 3) AddVBNode(n1, n2, vt);
205 case Op_AddVS: return new (C, 3) AddVSNode(n1, n2, vt);
206 case Op_AddVI: return new (C, 3) AddVINode(n1, n2, vt);
207 case Op_AddVL: return new (C, 3) AddVLNode(n1, n2, vt);
208 case Op_AddVF: return new (C, 3) AddVFNode(n1, n2, vt);
209 case Op_AddVD: return new (C, 3) AddVDNode(n1, n2, vt);
211 case Op_SubVB: return new (C, 3) SubVBNode(n1, n2, vt);
212 case Op_SubVS: return new (C, 3) SubVSNode(n1, n2, vt);
213 case Op_SubVI: return new (C, 3) SubVINode(n1, n2, vt);
214 case Op_SubVL: return new (C, 3) SubVLNode(n1, n2, vt);
215 case Op_SubVF: return new (C, 3) SubVFNode(n1, n2, vt);
216 case Op_SubVD: return new (C, 3) SubVDNode(n1, n2, vt);
218 case Op_MulVS: return new (C, 3) MulVSNode(n1, n2, vt);
219 case Op_MulVI: return new (C, 3) MulVINode(n1, n2, vt);
220 case Op_MulVF: return new (C, 3) MulVFNode(n1, n2, vt);
221 case Op_MulVD: return new (C, 3) MulVDNode(n1, n2, vt);
223 case Op_DivVF: return new (C, 3) DivVFNode(n1, n2, vt);
224 case Op_DivVD: return new (C, 3) DivVDNode(n1, n2, vt);
226 case Op_LShiftVB: return new (C, 3) LShiftVBNode(n1, n2, vt);
227 case Op_LShiftVS: return new (C, 3) LShiftVSNode(n1, n2, vt);
228 case Op_LShiftVI: return new (C, 3) LShiftVINode(n1, n2, vt);
229 case Op_LShiftVL: return new (C, 3) LShiftVLNode(n1, n2, vt);
231 case Op_RShiftVB: return new (C, 3) RShiftVBNode(n1, n2, vt);
232 case Op_RShiftVS: return new (C, 3) RShiftVSNode(n1, n2, vt);
233 case Op_RShiftVI: return new (C, 3) RShiftVINode(n1, n2, vt);
234 case Op_RShiftVL: return new (C, 3) RShiftVLNode(n1, n2, vt);
236 case Op_URShiftVB: return new (C, 3) URShiftVBNode(n1, n2, vt);
237 case Op_URShiftVS: return new (C, 3) URShiftVSNode(n1, n2, vt);
238 case Op_URShiftVI: return new (C, 3) URShiftVINode(n1, n2, vt);
239 case Op_URShiftVL: return new (C, 3) URShiftVLNode(n1, n2, vt);
241 case Op_AndV: return new (C, 3) AndVNode(n1, n2, vt);
242 case Op_OrV: return new (C, 3) OrVNode (n1, n2, vt);
243 case Op_XorV: return new (C, 3) XorVNode(n1, n2, vt);
244 }
245 ShouldNotReachHere();
246 return NULL;
248 }
250 // Scalar promotion
251 VectorNode* VectorNode::scalar2vector(Compile* C, Node* s, uint vlen, const Type* opd_t) {
252 BasicType bt = opd_t->array_element_basic_type();
253 const TypeVect* vt = opd_t->singleton() ? TypeVect::make(opd_t, vlen)
254 : TypeVect::make(bt, vlen);
255 switch (bt) {
256 case T_BOOLEAN:
257 case T_BYTE:
258 return new (C, 2) ReplicateBNode(s, vt);
259 case T_CHAR:
260 case T_SHORT:
261 return new (C, 2) ReplicateSNode(s, vt);
262 case T_INT:
263 return new (C, 2) ReplicateINode(s, vt);
264 case T_LONG:
265 return new (C, 2) ReplicateLNode(s, vt);
266 case T_FLOAT:
267 return new (C, 2) ReplicateFNode(s, vt);
268 case T_DOUBLE:
269 return new (C, 2) ReplicateDNode(s, vt);
270 }
271 ShouldNotReachHere();
272 return NULL;
273 }
275 // Return initial Pack node. Additional operands added with add_opd() calls.
276 PackNode* PackNode::make(Compile* C, Node* s, uint vlen, BasicType bt) {
277 const TypeVect* vt = TypeVect::make(bt, vlen);
278 switch (bt) {
279 case T_BOOLEAN:
280 case T_BYTE:
281 return new (C, vlen+1) PackBNode(s, vt);
282 case T_CHAR:
283 case T_SHORT:
284 return new (C, vlen+1) PackSNode(s, vt);
285 case T_INT:
286 return new (C, vlen+1) PackINode(s, vt);
287 case T_LONG:
288 return new (C, vlen+1) PackLNode(s, vt);
289 case T_FLOAT:
290 return new (C, vlen+1) PackFNode(s, vt);
291 case T_DOUBLE:
292 return new (C, vlen+1) PackDNode(s, vt);
293 }
294 ShouldNotReachHere();
295 return NULL;
296 }
298 // Create a binary tree form for Packs. [lo, hi) (half-open) range
299 Node* PackNode::binaryTreePack(Compile* C, int lo, int hi) {
300 int ct = hi - lo;
301 assert(is_power_of_2(ct), "power of 2");
302 if (ct == 2) {
303 PackNode* pk = PackNode::make(C, in(lo), 2, vect_type()->element_basic_type());
304 pk->add_opd(1, in(lo+1));
305 return pk;
307 } else {
308 int mid = lo + ct/2;
309 Node* n1 = binaryTreePack(C, lo, mid);
310 Node* n2 = binaryTreePack(C, mid, hi );
312 BasicType bt = vect_type()->element_basic_type();
313 switch (bt) {
314 case T_BOOLEAN:
315 case T_BYTE:
316 return new (C, 3) PackSNode(n1, n2, TypeVect::make(T_SHORT, 2));
317 case T_CHAR:
318 case T_SHORT:
319 return new (C, 3) PackINode(n1, n2, TypeVect::make(T_INT, 2));
320 case T_INT:
321 return new (C, 3) PackLNode(n1, n2, TypeVect::make(T_LONG, 2));
322 case T_LONG:
323 return new (C, 3) Pack2LNode(n1, n2, TypeVect::make(T_LONG, 2));
324 case T_FLOAT:
325 return new (C, 3) PackDNode(n1, n2, TypeVect::make(T_DOUBLE, 2));
326 case T_DOUBLE:
327 return new (C, 3) Pack2DNode(n1, n2, TypeVect::make(T_DOUBLE, 2));
328 }
329 ShouldNotReachHere();
330 }
331 return NULL;
332 }
334 // Return the vector version of a scalar load node.
335 LoadVectorNode* LoadVectorNode::make(Compile* C, int opc, Node* ctl, Node* mem,
336 Node* adr, const TypePtr* atyp, uint vlen, BasicType bt) {
337 const TypeVect* vt = TypeVect::make(bt, vlen);
338 return new (C, 3) LoadVectorNode(ctl, mem, adr, atyp, vt);
339 return NULL;
340 }
342 // Return the vector version of a scalar store node.
343 StoreVectorNode* StoreVectorNode::make(Compile* C, int opc, Node* ctl, Node* mem,
344 Node* adr, const TypePtr* atyp, Node* val,
345 uint vlen) {
346 return new (C, 4) StoreVectorNode(ctl, mem, adr, atyp, val);
347 }
349 // Extract a scalar element of vector.
350 Node* ExtractNode::make(Compile* C, Node* v, uint position, BasicType bt) {
351 assert((int)position < Matcher::max_vector_size(bt), "pos in range");
352 ConINode* pos = ConINode::make(C, (int)position);
353 switch (bt) {
354 case T_BOOLEAN:
355 return new (C, 3) ExtractUBNode(v, pos);
356 case T_BYTE:
357 return new (C, 3) ExtractBNode(v, pos);
358 case T_CHAR:
359 return new (C, 3) ExtractCNode(v, pos);
360 case T_SHORT:
361 return new (C, 3) ExtractSNode(v, pos);
362 case T_INT:
363 return new (C, 3) ExtractINode(v, pos);
364 case T_LONG:
365 return new (C, 3) ExtractLNode(v, pos);
366 case T_FLOAT:
367 return new (C, 3) ExtractFNode(v, pos);
368 case T_DOUBLE:
369 return new (C, 3) ExtractDNode(v, pos);
370 }
371 ShouldNotReachHere();
372 return NULL;
373 }