1 /*
2 * Copyright (c) 2021, 2022, 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 CPU_X86_MATCHER_X86_HPP
26 #define CPU_X86_MATCHER_X86_HPP
27
28 // Defined within class Matcher
29
30 // The ecx parameter to rep stosq for the ClearArray node is in words.
31 static const bool init_array_count_is_in_bytes = false;
32
33 // Whether this platform implements the scalable vector feature
34 static const bool implements_scalable_vector = false;
35
36 static constexpr bool supports_scalable_vector() {
37 return false;
38 }
39
40 // x86 supports misaligned vectors store/load.
41 static constexpr bool misaligned_vectors_ok() {
42 return true;
43 }
44
45 // Whether code generation need accurate ConvI2L types.
46 static const bool convi2l_type_required = true;
47
48 // Do the processor's shift instructions only use the low 5/6 bits
49 // of the count for 32/64 bit ints? If not we need to do the masking
50 // ourselves.
51 static const bool need_masked_shift_count = false;
52
53 // Does the CPU require late expand (see block.cpp for description of late expand)?
54 static const bool require_postalloc_expand = false;
55
56 // x86 supports generic vector operands: vec and legVec.
57 static const bool supports_generic_vector_operands = true;
58
59 static constexpr bool isSimpleConstant64(jlong value) {
60 // Will one (StoreL ConL) be cheaper than two (StoreI ConI)?.
61 //return value == (int) value; // Cf. storeImmL and immL32.
62
63 // Probably always true, even if a temp register is required.
64 #ifdef _LP64
65 return true;
66 #else
67 return false;
68 #endif
69 }
70
71 #ifdef _LP64
72 // No additional cost for CMOVL.
73 static constexpr int long_cmove_cost() { return 0; }
74 #else
75 // Needs 2 CMOV's for longs.
76 static constexpr int long_cmove_cost() { return 1; }
77 #endif
78
79 #ifdef _LP64
80 // No CMOVF/CMOVD with SSE2
81 static int float_cmove_cost() { return ConditionalMoveLimit; }
82 #else
83 // No CMOVF/CMOVD with SSE/SSE2
84 static int float_cmove_cost() { return (UseSSE>=1) ? ConditionalMoveLimit : 0; }
85 #endif
86
87 static bool narrow_oop_use_complex_address() {
88 NOT_LP64(ShouldNotCallThis();)
89 assert(UseCompressedOops, "only for compressed oops code");
90 return (LogMinObjAlignmentInBytes <= 3);
91 }
92
93 static bool narrow_klass_use_complex_address() {
94 NOT_LP64(ShouldNotCallThis();)
95 assert(UseCompressedClassPointers, "only for compressed klass code");
96 return (LogKlassAlignmentInBytes <= 3);
97 }
98
99 // Prefer ConN+DecodeN over ConP.
100 static bool const_oop_prefer_decode() {
101 NOT_LP64(ShouldNotCallThis();)
102 // Prefer ConN+DecodeN over ConP.
103 return true;
104 }
105
106 // Prefer ConP over ConNKlass+DecodeNKlass.
107 static bool const_klass_prefer_decode() {
108 NOT_LP64(ShouldNotCallThis();)
109 return false;
110 }
111
112 // Is it better to copy float constants, or load them directly from memory?
113 // Intel can load a float constant from a direct address, requiring no
114 // extra registers. Most RISCs will have to materialize an address into a
115 // register first, so they would do better to copy the constant from stack.
116 static const bool rematerialize_float_constants = true;
117
118 // If CPU can load and store mis-aligned doubles directly then no fixup is
119 // needed. Else we split the double into 2 integer pieces and move it
120 // piece-by-piece. Only happens when passing doubles into C code as the
121 // Java calling convention forces doubles to be aligned.
122 static const bool misaligned_doubles_ok = true;
123
124 // Advertise here if the CPU requires explicit rounding operations to implement strictfp mode.
125 #ifdef _LP64
126 static const bool strict_fp_requires_explicit_rounding = false;
127 #else
128 static const bool strict_fp_requires_explicit_rounding = true;
129 #endif
130
131 // Are floats converted to double when stored to stack during deoptimization?
132 // On x64 it is stored without conversion so we can use normal access.
133 // On x32 it is stored with conversion only when FPU is used for floats.
134 #ifdef _LP64
135 static constexpr bool float_in_double() {
136 return false;
137 }
138 #else
139 static bool float_in_double() {
140 return (UseSSE == 0);
141 }
142 #endif
143
144 // Do ints take an entire long register or just half?
145 #ifdef _LP64
146 static const bool int_in_long = true;
147 #else
148 static const bool int_in_long = false;
149 #endif
150
151
152 // Does the CPU supports vector variable shift instructions?
153 static bool supports_vector_variable_shifts(void) {
154 return (UseAVX >= 2);
155 }
156
157 // Does target support predicated operation emulation.
158 static bool supports_vector_predicate_op_emulation(int vopc, int vlen, BasicType bt) {
159 switch(vopc) {
160 case Op_LoadVectorGatherMasked:
161 return is_subword_type(bt) && VM_Version::supports_avx2();
162 default:
163 return false;
164 }
165 }
166
167 // Does the CPU supports vector variable rotate instructions?
168 static constexpr bool supports_vector_variable_rotates(void) {
169 return true;
170 }
171
172 // Does the CPU supports vector constant rotate instructions?
173 static constexpr bool supports_vector_constant_rotates(int shift) {
174 return -0x80 <= shift && shift < 0x80;
175 }
176
177 // Does the CPU supports vector unsigned comparison instructions?
178 static constexpr bool supports_vector_comparison_unsigned(int vlen, BasicType bt) {
179 return true;
180 }
181
182 // Some microarchitectures have mask registers used on vectors
183 static bool has_predicated_vectors(void) {
184 return VM_Version::supports_evex();
185 }
186
187 // true means we have fast l2f conversion
188 // false means that conversion is done by runtime call
189 static constexpr bool convL2FSupported(void) {
190 return true;
191 }
192
193 // Implements a variant of EncodeISOArrayNode that encode ASCII only
194 static const bool supports_encode_ascii_array = true;
195
196 // Without predicated input, an all-one vector is needed for the alltrue vector test
197 static constexpr bool vectortest_needs_second_argument(bool is_alltrue, bool is_predicate) {
198 return is_alltrue && !is_predicate;
199 }
200
201 // BoolTest mask for vector test intrinsics
202 static constexpr BoolTest::mask vectortest_mask(bool is_alltrue, bool is_predicate, int vlen) {
203 if (!is_alltrue) {
204 return BoolTest::ne;
205 }
206 if (!is_predicate) {
207 return BoolTest::lt;
208 }
209 if ((vlen == 8 && !VM_Version::supports_avx512dq()) || vlen < 8) {
210 return BoolTest::eq;
211 }
212 return BoolTest::lt;
213 }
214
215 // Returns pre-selection estimated size of a vector operation.
216 // Currently, it's a rudimentary heuristic based on emitted code size for complex
217 // IR nodes used by unroll policy. Idea is to constrain unrolling factor and prevent
218 // generating bloated loop bodies.
219 static int vector_op_pre_select_sz_estimate(int vopc, BasicType ety, int vlen) {
220 switch(vopc) {
221 default:
222 return 0;
223 case Op_MulVB:
224 return 7;
225 case Op_MulVL:
226 return VM_Version::supports_avx512vldq() ? 0 : 6;
227 case Op_LoadVectorGather:
228 case Op_LoadVectorGatherMasked:
229 return is_subword_type(ety) ? 50 : 0;
230 case Op_VectorCastF2X: // fall through
231 case Op_VectorCastD2X:
232 return is_floating_point_type(ety) ? 0 : (is_subword_type(ety) ? 35 : 30);
233 case Op_CountTrailingZerosV:
234 case Op_CountLeadingZerosV:
235 return VM_Version::supports_avx512cd() && (ety == T_INT || ety == T_LONG) ? 0 : 40;
236 case Op_PopCountVI:
237 if (is_subword_type(ety)) {
238 return VM_Version::supports_avx512_bitalg() ? 0 : 50;
239 } else {
240 assert(ety == T_INT, "sanity"); // for documentation purposes
241 return VM_Version::supports_avx512_vpopcntdq() ? 0 : 50;
242 }
243 case Op_PopCountVL:
244 return VM_Version::supports_avx512_vpopcntdq() ? 0 : 40;
245 case Op_ReverseV:
246 return VM_Version::supports_gfni() ? 0 : 30;
247 case Op_RoundVF: // fall through
248 case Op_RoundVD:
249 return 30;
250 }
251 }
252
253 // Returns pre-selection estimated size of a scalar operation.
254 static int scalar_op_pre_select_sz_estimate(int vopc, BasicType ety) {
255 switch(vopc) {
256 default: return 0;
257 case Op_RoundF: // fall through
258 case Op_RoundD: {
259 return 30;
260 }
261 }
262 }
263
264 // Is SIMD sort supported for this CPU?
265 static bool supports_simd_sort(BasicType bt) {
266 if (VM_Version::supports_avx512dq()) {
267 return true;
268 }
269 else if (VM_Version::supports_avx2() && !is_double_word_type(bt)) {
270 return true;
271 }
272 else {
273 return false;
274 }
275 }
276
277 #endif // CPU_X86_MATCHER_X86_HPP