125
126 void evmovdqu(BasicType type, KRegister kmask, XMMRegister dst, Address src, int vector_len);
127 void evmovdqu(BasicType type, KRegister kmask, Address dst, XMMRegister src, int vector_len);
128
129 // extract
130 void extract(BasicType typ, Register dst, XMMRegister src, int idx);
131 XMMRegister get_lane(BasicType typ, XMMRegister dst, XMMRegister src, int elemindex);
132 void get_elem(BasicType typ, Register dst, XMMRegister src, int elemindex);
133 void get_elem(BasicType typ, XMMRegister dst, XMMRegister src, int elemindex, Register tmp = noreg, XMMRegister vtmp = xnoreg);
134
135 // vector test
136 void vectortest(int bt, int vlen, XMMRegister src1, XMMRegister src2,
137 XMMRegister vtmp1 = xnoreg, XMMRegister vtmp2 = xnoreg, KRegister mask = knoreg);
138
139 // blend
140 void evpcmp(BasicType typ, KRegister kdmask, KRegister ksmask, XMMRegister src1, AddressLiteral adr, int comparison, int vector_len, Register scratch = rscratch1);
141 void evpcmp(BasicType typ, KRegister kdmask, KRegister ksmask, XMMRegister src1, XMMRegister src2, int comparison, int vector_len);
142 void evpblend(BasicType typ, XMMRegister dst, KRegister kmask, XMMRegister src1, XMMRegister src2, bool merge, int vector_len);
143
144 void load_vector_mask(XMMRegister dst, XMMRegister src, int vlen_in_bytes, BasicType elem_bt, bool is_legacy);
145 void load_iota_indices(XMMRegister dst, Register scratch, int vlen_in_bytes);
146
147 // vector compare
148 void vpcmpu(BasicType typ, XMMRegister dst, XMMRegister src1, XMMRegister src2, ComparisonPredicate comparison, int vlen_in_bytes,
149 XMMRegister vtmp1, XMMRegister vtmp2, Register scratch);
150 void vpcmpu32(BasicType typ, XMMRegister dst, XMMRegister src1, XMMRegister src2, ComparisonPredicate comparison, int vlen_in_bytes,
151 XMMRegister vtmp1, XMMRegister vtmp2, XMMRegister vtmp3, Register scratch);
152
153 // Reductions for vectors of bytes, shorts, ints, longs, floats, and doubles.
154
155 // dst = src1 reduce(op, src2) using vtmp as temps
156 void reduceI(int opcode, int vlen, Register dst, Register src1, XMMRegister src2, XMMRegister vtmp1, XMMRegister vtmp2);
157 #ifdef _LP64
158 void reduceL(int opcode, int vlen, Register dst, Register src1, XMMRegister src2, XMMRegister vtmp1, XMMRegister vtmp2);
159 void genmask(KRegister dst, Register len, Register temp);
160 #endif // _LP64
161
162 // dst = reduce(op, src2) using vtmp as temps
163 void reduce_fp(int opcode, int vlen,
164 XMMRegister dst, XMMRegister src,
205 void reduce8L(int opcode, Register dst, Register src1, XMMRegister src2, XMMRegister vtmp1, XMMRegister vtmp2);
206 #endif // _LP64
207
208 // Float Reduction
209 void reduce2F (int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp);
210 void reduce4F (int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp);
211 void reduce8F (int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp1, XMMRegister vtmp2);
212 void reduce16F(int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp1, XMMRegister vtmp2);
213
214 // Double Reduction
215 void reduce2D(int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp);
216 void reduce4D(int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp1, XMMRegister vtmp2);
217 void reduce8D(int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp1, XMMRegister vtmp2);
218
219 // Base reduction instruction
220 void reduce_operation_128(BasicType typ, int opcode, XMMRegister dst, XMMRegister src);
221 void reduce_operation_256(BasicType typ, int opcode, XMMRegister dst, XMMRegister src1, XMMRegister src2);
222
223 public:
224 #ifdef _LP64
225 void vector_mask_operation(int opc, Register dst, XMMRegister mask, XMMRegister xtmp, Register tmp,
226 KRegister ktmp, int masklen, int vec_enc);
227
228 void vector_mask_operation(int opc, Register dst, XMMRegister mask, XMMRegister xtmp, XMMRegister xtmp1,
229 Register tmp, int masklen, int vec_enc);
230 #endif
231 void string_indexof_char(Register str1, Register cnt1, Register ch, Register result,
232 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp);
233
234 void stringL_indexof_char(Register str1, Register cnt1, Register ch, Register result,
235 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp);
236
237 // IndexOf strings.
238 // Small strings are loaded through stack if they cross page boundary.
239 void string_indexof(Register str1, Register str2,
240 Register cnt1, Register cnt2,
241 int int_cnt2, Register result,
242 XMMRegister vec, Register tmp,
243 int ae);
244
245 // IndexOf for constant substrings with size >= 8 elements
246 // which don't need to be loaded through stack.
247 void string_indexofC8(Register str1, Register str2,
248 Register cnt1, Register cnt2,
249 int int_cnt2, Register result,
256 // helper function for string_compare
257 void load_next_elements(Register elem1, Register elem2, Register str1, Register str2,
258 Address::ScaleFactor scale, Address::ScaleFactor scale1,
259 Address::ScaleFactor scale2, Register index, int ae);
260 // Compare strings.
261 void string_compare(Register str1, Register str2,
262 Register cnt1, Register cnt2, Register result,
263 XMMRegister vec1, int ae, KRegister mask = knoreg);
264
265 // Search for Non-ASCII character (Negative byte value) in a byte array,
266 // return true if it has any and false otherwise.
267 void has_negatives(Register ary1, Register len,
268 Register result, Register tmp1,
269 XMMRegister vec1, XMMRegister vec2, KRegister mask1 = knoreg, KRegister mask2 = knoreg);
270
271 // Compare char[] or byte[] arrays.
272 void arrays_equals(bool is_array_equ, Register ary1, Register ary2,
273 Register limit, Register result, Register chr,
274 XMMRegister vec1, XMMRegister vec2, bool is_char, KRegister mask = knoreg);
275
276 #endif // CPU_X86_C2_MACROASSEMBLER_X86_HPP
|
125
126 void evmovdqu(BasicType type, KRegister kmask, XMMRegister dst, Address src, int vector_len);
127 void evmovdqu(BasicType type, KRegister kmask, Address dst, XMMRegister src, int vector_len);
128
129 // extract
130 void extract(BasicType typ, Register dst, XMMRegister src, int idx);
131 XMMRegister get_lane(BasicType typ, XMMRegister dst, XMMRegister src, int elemindex);
132 void get_elem(BasicType typ, Register dst, XMMRegister src, int elemindex);
133 void get_elem(BasicType typ, XMMRegister dst, XMMRegister src, int elemindex, Register tmp = noreg, XMMRegister vtmp = xnoreg);
134
135 // vector test
136 void vectortest(int bt, int vlen, XMMRegister src1, XMMRegister src2,
137 XMMRegister vtmp1 = xnoreg, XMMRegister vtmp2 = xnoreg, KRegister mask = knoreg);
138
139 // blend
140 void evpcmp(BasicType typ, KRegister kdmask, KRegister ksmask, XMMRegister src1, AddressLiteral adr, int comparison, int vector_len, Register scratch = rscratch1);
141 void evpcmp(BasicType typ, KRegister kdmask, KRegister ksmask, XMMRegister src1, XMMRegister src2, int comparison, int vector_len);
142 void evpblend(BasicType typ, XMMRegister dst, KRegister kmask, XMMRegister src1, XMMRegister src2, bool merge, int vector_len);
143
144 void load_vector_mask(XMMRegister dst, XMMRegister src, int vlen_in_bytes, BasicType elem_bt, bool is_legacy);
145 void load_vector_mask(KRegister dst, XMMRegister src, XMMRegister xtmp, Register tmp, bool novlbwdq, int vlen_enc);
146
147 void load_iota_indices(XMMRegister dst, Register scratch, int vlen_in_bytes);
148
149 // vector compare
150 void vpcmpu(BasicType typ, XMMRegister dst, XMMRegister src1, XMMRegister src2, ComparisonPredicate comparison, int vlen_in_bytes,
151 XMMRegister vtmp1, XMMRegister vtmp2, Register scratch);
152 void vpcmpu32(BasicType typ, XMMRegister dst, XMMRegister src1, XMMRegister src2, ComparisonPredicate comparison, int vlen_in_bytes,
153 XMMRegister vtmp1, XMMRegister vtmp2, XMMRegister vtmp3, Register scratch);
154
155 // Reductions for vectors of bytes, shorts, ints, longs, floats, and doubles.
156
157 // dst = src1 reduce(op, src2) using vtmp as temps
158 void reduceI(int opcode, int vlen, Register dst, Register src1, XMMRegister src2, XMMRegister vtmp1, XMMRegister vtmp2);
159 #ifdef _LP64
160 void reduceL(int opcode, int vlen, Register dst, Register src1, XMMRegister src2, XMMRegister vtmp1, XMMRegister vtmp2);
161 void genmask(KRegister dst, Register len, Register temp);
162 #endif // _LP64
163
164 // dst = reduce(op, src2) using vtmp as temps
165 void reduce_fp(int opcode, int vlen,
166 XMMRegister dst, XMMRegister src,
207 void reduce8L(int opcode, Register dst, Register src1, XMMRegister src2, XMMRegister vtmp1, XMMRegister vtmp2);
208 #endif // _LP64
209
210 // Float Reduction
211 void reduce2F (int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp);
212 void reduce4F (int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp);
213 void reduce8F (int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp1, XMMRegister vtmp2);
214 void reduce16F(int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp1, XMMRegister vtmp2);
215
216 // Double Reduction
217 void reduce2D(int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp);
218 void reduce4D(int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp1, XMMRegister vtmp2);
219 void reduce8D(int opcode, XMMRegister dst, XMMRegister src, XMMRegister vtmp1, XMMRegister vtmp2);
220
221 // Base reduction instruction
222 void reduce_operation_128(BasicType typ, int opcode, XMMRegister dst, XMMRegister src);
223 void reduce_operation_256(BasicType typ, int opcode, XMMRegister dst, XMMRegister src1, XMMRegister src2);
224
225 public:
226 #ifdef _LP64
227 void vector_mask_operation(int opc, Register dst, KRegister mask, Register tmp, int masklen, int masksize, int vec_enc);
228
229 void vector_mask_operation(int opc, Register dst, XMMRegister mask, XMMRegister xtmp, XMMRegister xtmp1,
230 Register tmp, int masklen, int masksize, int vec_enc);
231 #endif
232 void string_indexof_char(Register str1, Register cnt1, Register ch, Register result,
233 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp);
234
235 void stringL_indexof_char(Register str1, Register cnt1, Register ch, Register result,
236 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp);
237
238 // IndexOf strings.
239 // Small strings are loaded through stack if they cross page boundary.
240 void string_indexof(Register str1, Register str2,
241 Register cnt1, Register cnt2,
242 int int_cnt2, Register result,
243 XMMRegister vec, Register tmp,
244 int ae);
245
246 // IndexOf for constant substrings with size >= 8 elements
247 // which don't need to be loaded through stack.
248 void string_indexofC8(Register str1, Register str2,
249 Register cnt1, Register cnt2,
250 int int_cnt2, Register result,
257 // helper function for string_compare
258 void load_next_elements(Register elem1, Register elem2, Register str1, Register str2,
259 Address::ScaleFactor scale, Address::ScaleFactor scale1,
260 Address::ScaleFactor scale2, Register index, int ae);
261 // Compare strings.
262 void string_compare(Register str1, Register str2,
263 Register cnt1, Register cnt2, Register result,
264 XMMRegister vec1, int ae, KRegister mask = knoreg);
265
266 // Search for Non-ASCII character (Negative byte value) in a byte array,
267 // return true if it has any and false otherwise.
268 void has_negatives(Register ary1, Register len,
269 Register result, Register tmp1,
270 XMMRegister vec1, XMMRegister vec2, KRegister mask1 = knoreg, KRegister mask2 = knoreg);
271
272 // Compare char[] or byte[] arrays.
273 void arrays_equals(bool is_array_equ, Register ary1, Register ary2,
274 Register limit, Register result, Register chr,
275 XMMRegister vec1, XMMRegister vec2, bool is_char, KRegister mask = knoreg);
276
277
278 void evmasked_op(int ideal_opc, BasicType eType, KRegister mask,
279 XMMRegister dst, XMMRegister src1, XMMRegister src2,
280 bool merge, int vlen_enc, bool is_varshift = false);
281
282 void evmasked_op(int ideal_opc, BasicType eType, KRegister mask,
283 XMMRegister dst, XMMRegister src1, Address src2,
284 bool merge, int vlen_enc);
285
286 void evmasked_op(int ideal_opc, BasicType eType, KRegister mask, XMMRegister dst,
287 XMMRegister src1, int imm8, bool merge, int vlen_enc);
288
289 void masked_op(int ideal_opc, int mask_len, KRegister dst,
290 KRegister src1, KRegister src2);
291 #endif // CPU_X86_C2_MACROASSEMBLER_X86_HPP
|