Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(210)

Side by Side Diff: src/a64/codegen-a64.cc

Issue 144963003: A64: add missing files. (Closed) Base URL: https://v8.googlecode.com/svn/branches/experimental/a64
Patch Set: Created 6 years, 11 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View unified diff | Download patch | Annotate | Revision Log
« no previous file with comments | « src/a64/codegen-a64.h ('k') | src/a64/constants-a64.h » ('j') | no next file with comments »
Toggle Intra-line Diffs ('i') | Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
OLDNEW
(Empty)
1 // Copyright 2013 the V8 project authors. All rights reserved.
2 // Redistribution and use in source and binary forms, with or without
3 // modification, are permitted provided that the following conditions are
4 // met:
5 //
6 // * Redistributions of source code must retain the above copyright
7 // notice, this list of conditions and the following disclaimer.
8 // * Redistributions in binary form must reproduce the above
9 // copyright notice, this list of conditions and the following
10 // disclaimer in the documentation and/or other materials provided
11 // with the distribution.
12 // * Neither the name of Google Inc. nor the names of its
13 // contributors may be used to endorse or promote products derived
14 // from this software without specific prior written permission.
15 //
16 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
17 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
18 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
19 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
20 // OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
21 // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
22 // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
26 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27
28 #include "v8.h"
29
30 #if defined(V8_TARGET_ARCH_A64)
31
32 #include "codegen.h"
33 #include "macro-assembler.h"
34 #include "simulator-a64.h"
35
36 namespace v8 {
37 namespace internal {
38
39 #define __ ACCESS_MASM(masm)
40
41 UnaryMathFunction CreateTranscendentalFunction(TranscendentalCache::Type type) {
42 switch (type) {
43 case TranscendentalCache::SIN: return &sin;
44 case TranscendentalCache::COS: return &cos;
45 case TranscendentalCache::TAN: return &tan;
46 case TranscendentalCache::LOG: return &log;
47 default: UNIMPLEMENTED();
48 }
49 return NULL;
50 }
51
52
53 #if defined(USE_SIMULATOR)
54 byte* fast_exp_a64_machine_code = NULL;
55 double fast_exp_simulator(double x) {
56 Simulator * simulator = Simulator::current(Isolate::Current());
57 return simulator->CallDouble(fast_exp_a64_machine_code,
58 Simulator::CallArgument(x),
59 Simulator::CallArgument::End());
60 }
61 #endif
62
63
64 UnaryMathFunction CreateExpFunction() {
65 if (!FLAG_fast_math) return &exp;
66
67 // Use the Math.exp implemetation in MathExpGenerator::EmitMathExp() to create
68 // an AAPCS64-compliant exp() function. This will be faster than the C
69 // library's exp() function, but probably less accurate.
70 size_t actual_size;
71 byte* buffer = static_cast<byte*>(OS::Allocate(1 * KB, &actual_size, true));
72 if (buffer == NULL) return &exp;
73
74 ExternalReference::InitializeMathExpData();
75 MacroAssembler masm(NULL, buffer, static_cast<int>(actual_size));
76 masm.SetStackPointer(csp);
77
78 // The argument will be in d0 on entry.
79 DoubleRegister input = d0;
80 // Use other caller-saved registers for all other values.
81 DoubleRegister result = d1;
82 DoubleRegister double_temp1 = d2;
83 DoubleRegister double_temp2 = d3;
84 Register temp1 = x10;
85 Register temp2 = x11;
86 Register temp3 = x12;
87
88 MathExpGenerator::EmitMathExp(&masm, input, result,
89 double_temp1, double_temp2,
90 temp1, temp2, temp3);
91 // Move the result to the return register.
92 masm.Fmov(d0, result);
93 masm.Ret();
94
95 CodeDesc desc;
96 masm.GetCode(&desc);
97 ASSERT(!RelocInfo::RequiresRelocation(desc));
98
99 CPU::FlushICache(buffer, actual_size);
100 OS::ProtectCode(buffer, actual_size);
101
102 #if !defined(USE_SIMULATOR)
103 return FUNCTION_CAST<UnaryMathFunction>(buffer);
104 #else
105 fast_exp_a64_machine_code = buffer;
106 return &fast_exp_simulator;
107 #endif
108 }
109
110
111 UnaryMathFunction CreateSqrtFunction() {
112 return &sqrt;
113 }
114
115 // -------------------------------------------------------------------------
116 // Platform-specific RuntimeCallHelper functions.
117
118 void StubRuntimeCallHelper::BeforeCall(MacroAssembler* masm) const {
119 masm->EnterFrame(StackFrame::INTERNAL);
120 ASSERT(!masm->has_frame());
121 masm->set_has_frame(true);
122 }
123
124
125 void StubRuntimeCallHelper::AfterCall(MacroAssembler* masm) const {
126 masm->LeaveFrame(StackFrame::INTERNAL);
127 ASSERT(masm->has_frame());
128 masm->set_has_frame(false);
129 }
130
131
132 // -------------------------------------------------------------------------
133 // Code generators
134
135 void ElementsTransitionGenerator::GenerateMapChangeElementsTransition(
136 MacroAssembler* masm, AllocationSiteMode mode,
137 Label* allocation_site_info_found) {
138 // ----------- S t a t e -------------
139 // -- x2 : receiver
140 // -- x3 : target map
141 // -----------------------------------
142 Register receiver = x2;
143 Register map = x3;
144
145 if (mode == TRACK_ALLOCATION_SITE) {
146 ASSERT(allocation_site_info_found != NULL);
147 __ TestJSArrayForAllocationSiteInfo(receiver, x10, x11);
148 __ B(eq, allocation_site_info_found);
149 }
150
151 // Set transitioned map.
152 __ Str(map, FieldMemOperand(receiver, HeapObject::kMapOffset));
153 __ RecordWriteField(receiver,
154 HeapObject::kMapOffset,
155 map,
156 x10,
157 kLRHasNotBeenSaved,
158 kDontSaveFPRegs,
159 EMIT_REMEMBERED_SET,
160 OMIT_SMI_CHECK);
161 }
162
163
164 void ElementsTransitionGenerator::GenerateSmiToDouble(
165 MacroAssembler* masm, AllocationSiteMode mode, Label* fail) {
166 ASM_LOCATION("ElementsTransitionGenerator::GenerateSmiToDouble");
167 // ----------- S t a t e -------------
168 // -- lr : return address
169 // -- x0 : value
170 // -- x1 : key
171 // -- x2 : receiver
172 // -- x3 : target map, scratch for subsequent call
173 // -----------------------------------
174 Register receiver = x2;
175 Register target_map = x3;
176
177 Label gc_required, only_change_map;
178
179 if (mode == TRACK_ALLOCATION_SITE) {
180 __ TestJSArrayForAllocationSiteInfo(receiver, x10, x11);
181 __ B(eq, fail);
182 }
183
184 // Check for empty arrays, which only require a map transition and no changes
185 // to the backing store.
186 Register elements = x4;
187 __ Ldr(elements, FieldMemOperand(receiver, JSObject::kElementsOffset));
188 __ JumpIfRoot(elements, Heap::kEmptyFixedArrayRootIndex, &only_change_map);
189
190 __ Push(lr);
191 Register length = x5;
192 __ Ldrsw(length, UntagSmiFieldMemOperand(elements,
193 FixedArray::kLengthOffset));
194
195 // Allocate new FixedDoubleArray.
196 Register array_size = x6;
197 Register array = x7;
198 __ Lsl(array_size, length, kDoubleSizeLog2);
199 __ Add(array_size, array_size, FixedDoubleArray::kHeaderSize);
200 __ Allocate(array_size, array, x10, x11, &gc_required, DOUBLE_ALIGNMENT);
201 // Register array is non-tagged heap object.
202
203 // Set the destination FixedDoubleArray's length and map.
204 Register map_root = x6;
205 __ LoadRoot(map_root, Heap::kFixedDoubleArrayMapRootIndex);
206 __ SmiTag(x11, length);
207 __ Str(x11, MemOperand(array, FixedDoubleArray::kLengthOffset));
208 __ Str(map_root, MemOperand(array, HeapObject::kMapOffset));
209
210 __ Str(target_map, FieldMemOperand(receiver, HeapObject::kMapOffset));
211 __ RecordWriteField(receiver, HeapObject::kMapOffset, target_map, x6,
212 kLRHasBeenSaved, kDontSaveFPRegs, OMIT_REMEMBERED_SET,
213 OMIT_SMI_CHECK, EXPECT_PREGENERATED);
214
215 // Replace receiver's backing store with newly created FixedDoubleArray.
216 __ Add(x10, array, kHeapObjectTag);
217 __ Str(x10, FieldMemOperand(receiver, JSObject::kElementsOffset));
218 __ RecordWriteField(receiver, JSObject::kElementsOffset, x10,
219 x6, kLRHasBeenSaved, kDontSaveFPRegs,
220 EMIT_REMEMBERED_SET, OMIT_SMI_CHECK, EXPECT_PREGENERATED);
221
222 // Prepare for conversion loop.
223 Register src_elements = x10;
224 Register dst_elements = x11;
225 Register dst_end = x12;
226 __ Add(src_elements, elements, FixedArray::kHeaderSize - kHeapObjectTag);
227 __ Add(dst_elements, array, FixedDoubleArray::kHeaderSize);
228 __ Add(dst_end, dst_elements, Operand(length, LSL, kDoubleSizeLog2));
229
230 FPRegister nan_d = d1;
231 __ Fmov(nan_d, rawbits_to_double(kHoleNanInt64));
232
233 Label entry, done;
234 __ B(&entry);
235
236 __ Bind(&only_change_map);
237 __ Str(target_map, FieldMemOperand(receiver, HeapObject::kMapOffset));
238 __ RecordWriteField(receiver, HeapObject::kMapOffset, target_map, x6,
239 kLRHasNotBeenSaved, kDontSaveFPRegs, OMIT_REMEMBERED_SET,
240 OMIT_SMI_CHECK, EXPECT_PREGENERATED);
241 __ B(&done);
242
243 // Call into runtime if GC is required.
244 __ Bind(&gc_required);
245 __ Pop(lr);
246 __ B(fail);
247
248 // Iterate over the array, copying and coverting smis to doubles. If an
249 // element is non-smi, write a hole to the destination.
250 {
251 Label loop;
252 __ Bind(&loop);
253 __ Ldr(x13, MemOperand(src_elements, kPointerSize, PostIndex));
254 __ SmiUntagToDouble(d0, x13, kSpeculativeUntag);
255 __ Tst(x13, kSmiTagMask);
256 __ Fcsel(d0, d0, nan_d, eq);
257 __ Str(d0, MemOperand(dst_elements, kDoubleSize, PostIndex));
258
259 __ Bind(&entry);
260 __ Cmp(dst_elements, dst_end);
261 __ B(lt, &loop);
262 }
263
264 __ Pop(lr);
265 __ Bind(&done);
266 }
267
268
269 void ElementsTransitionGenerator::GenerateDoubleToObject(
270 MacroAssembler* masm, AllocationSiteMode mode, Label* fail) {
271 ASM_LOCATION("ElementsTransitionGenerator::GenerateDoubleToObject");
272 // ----------- S t a t e -------------
273 // -- x0 : value
274 // -- x1 : key
275 // -- x2 : receiver
276 // -- lr : return address
277 // -- x3 : target map, scratch for subsequent call
278 // -- x4 : scratch (elements)
279 // -----------------------------------
280 Register value = x0;
281 Register key = x1;
282 Register receiver = x2;
283 Register target_map = x3;
284
285 if (mode == TRACK_ALLOCATION_SITE) {
286 __ TestJSArrayForAllocationSiteInfo(receiver, x10, x11);
287 __ B(eq, fail);
288 }
289
290 // Check for empty arrays, which only require a map transition and no changes
291 // to the backing store.
292 Label only_change_map;
293 Register elements = x4;
294 __ Ldr(elements, FieldMemOperand(receiver, JSObject::kElementsOffset));
295 __ JumpIfRoot(elements, Heap::kEmptyFixedArrayRootIndex, &only_change_map);
296
297 __ Push(lr);
298 // TODO(all): These registers may not need to be pushed. Examine
299 // RecordWriteStub and check whether it's needed.
300 __ Push(target_map, receiver, key, value);
301 Register length = x5;
302 __ Ldrsw(length, UntagSmiFieldMemOperand(elements,
303 FixedArray::kLengthOffset));
304
305 // Allocate new FixedArray.
306 Register array_size = x6;
307 Register array = x7;
308 Label gc_required;
309 __ Mov(array_size, FixedDoubleArray::kHeaderSize);
310 __ Add(array_size, array_size, Operand(length, LSL, kPointerSizeLog2));
311 __ Allocate(array_size, array, x10, x11, &gc_required, NO_ALLOCATION_FLAGS);
312
313 // Set destination FixedDoubleArray's length and map.
314 Register map_root = x6;
315 __ LoadRoot(map_root, Heap::kFixedArrayMapRootIndex);
316 __ SmiTag(x11, length);
317 __ Str(x11, MemOperand(array, FixedDoubleArray::kLengthOffset));
318 __ Str(map_root, MemOperand(array, HeapObject::kMapOffset));
319
320 // Prepare for conversion loop.
321 Register src_elements = x10;
322 Register dst_elements = x11;
323 Register dst_end = x12;
324 __ Add(src_elements, elements,
325 FixedDoubleArray::kHeaderSize - kHeapObjectTag);
326 __ Add(dst_elements, array, FixedArray::kHeaderSize);
327 __ Add(array, array, kHeapObjectTag);
328 __ Add(dst_end, dst_elements, Operand(length, LSL, kPointerSizeLog2));
329
330 Register the_hole = x14;
331 Register heap_num_map = x15;
332 __ LoadRoot(the_hole, Heap::kTheHoleValueRootIndex);
333 __ LoadRoot(heap_num_map, Heap::kHeapNumberMapRootIndex);
334
335 Label entry;
336 __ B(&entry);
337
338 // Call into runtime if GC is required.
339 __ Bind(&gc_required);
340 __ Pop(value, key, receiver, target_map);
341 __ Pop(lr);
342 __ B(fail);
343
344 {
345 Label loop, convert_hole;
346 __ Bind(&loop);
347 __ Ldr(x13, MemOperand(src_elements, kPointerSize, PostIndex));
348 __ Cmp(x13, kHoleNanInt64);
349 __ B(eq, &convert_hole);
350
351 // Non-hole double, copy value into a heap number.
352 Register heap_num = x5;
353 __ AllocateHeapNumber(heap_num, &gc_required, x6, x4, heap_num_map);
354 __ Str(x13, FieldMemOperand(heap_num, HeapNumber::kValueOffset));
355 __ Mov(x13, dst_elements);
356 __ Str(heap_num, MemOperand(dst_elements, kPointerSize, PostIndex));
357 __ RecordWrite(array, x13, heap_num, kLRHasBeenSaved, kDontSaveFPRegs,
358 EMIT_REMEMBERED_SET, OMIT_SMI_CHECK, EXPECT_PREGENERATED);
359
360 __ B(&entry);
361
362 // Replace the-hole NaN with the-hole pointer.
363 __ Bind(&convert_hole);
364 __ Str(the_hole, MemOperand(dst_elements, kPointerSize, PostIndex));
365
366 __ Bind(&entry);
367 __ Cmp(dst_elements, dst_end);
368 __ B(lt, &loop);
369 }
370
371 __ Pop(value, key, receiver, target_map);
372 // Replace receiver's backing store with newly created and filled FixedArray.
373 __ Str(array, FieldMemOperand(receiver, JSObject::kElementsOffset));
374 __ RecordWriteField(receiver, JSObject::kElementsOffset, array, x13,
375 kLRHasBeenSaved, kDontSaveFPRegs, EMIT_REMEMBERED_SET,
376 OMIT_SMI_CHECK, EXPECT_PREGENERATED);
377 __ Pop(lr);
378
379 __ Bind(&only_change_map);
380 __ Str(target_map, FieldMemOperand(receiver, HeapObject::kMapOffset));
381 __ RecordWriteField(receiver, HeapObject::kMapOffset, target_map, x13,
382 kLRHasNotBeenSaved, kDontSaveFPRegs, OMIT_REMEMBERED_SET,
383 OMIT_SMI_CHECK, EXPECT_PREGENERATED);
384 }
385
386
387 bool Code::IsYoungSequence(byte* sequence) {
388 return MacroAssembler::IsYoungSequence(sequence);
389 }
390
391
392 void Code::GetCodeAgeAndParity(byte* sequence, Age* age,
393 MarkingParity* parity) {
394 if (IsYoungSequence(sequence)) {
395 *age = kNoAge;
396 *parity = NO_MARKING_PARITY;
397 } else {
398 byte* target = sequence + kCodeAgeStubEntryOffset;
399 Code* stub = GetCodeFromTargetAddress(Memory::Address_at(target));
400 GetCodeAgeAndParity(stub, age, parity);
401 }
402 }
403
404
405 void Code::PatchPlatformCodeAge(byte* sequence,
406 Code::Age age,
407 MarkingParity parity) {
408 PatchingAssembler patcher(sequence, kCodeAgeSequenceSize / kInstructionSize);
409 if (age == kNoAge) {
410 MacroAssembler::EmitFrameSetupForCodeAgePatching(&patcher);
411 } else {
412 Code * stub = GetCodeAgeStub(age, parity);
413 MacroAssembler::EmitCodeAgeSequence(&patcher, stub);
414 }
415 }
416
417
418 void StringCharLoadGenerator::Generate(MacroAssembler* masm,
419 Register string,
420 Register index,
421 Register result,
422 Label* call_runtime) {
423 // Fetch the instance type of the receiver into result register.
424 __ Ldr(result, FieldMemOperand(string, HeapObject::kMapOffset));
425 __ Ldrb(result, FieldMemOperand(result, Map::kInstanceTypeOffset));
426
427 // We need special handling for indirect strings.
428 Label check_sequential;
429 __ TestAndBranchIfAllClear(result, kIsIndirectStringMask, &check_sequential);
430
431 // Dispatch on the indirect string shape: slice or cons.
432 Label cons_string;
433 __ TestAndBranchIfAllClear(result, kSlicedNotConsMask, &cons_string);
434
435 // Handle slices.
436 Label indirect_string_loaded;
437 __ Ldrsw(result,
438 UntagSmiFieldMemOperand(string, SlicedString::kOffsetOffset));
439 __ Ldr(string, FieldMemOperand(string, SlicedString::kParentOffset));
440 __ Add(index, index, result);
441 __ B(&indirect_string_loaded);
442
443 // Handle cons strings.
444 // Check whether the right hand side is the empty string (i.e. if
445 // this is really a flat string in a cons string). If that is not
446 // the case we would rather go to the runtime system now to flatten
447 // the string.
448 __ Bind(&cons_string);
449 __ Ldr(result, FieldMemOperand(string, ConsString::kSecondOffset));
450 __ JumpIfNotRoot(result, Heap::kempty_stringRootIndex, call_runtime);
451 // Get the first of the two strings and load its instance type.
452 __ Ldr(string, FieldMemOperand(string, ConsString::kFirstOffset));
453
454 __ Bind(&indirect_string_loaded);
455 __ Ldr(result, FieldMemOperand(string, HeapObject::kMapOffset));
456 __ Ldrb(result, FieldMemOperand(result, Map::kInstanceTypeOffset));
457
458 // Distinguish sequential and external strings. Only these two string
459 // representations can reach here (slices and flat cons strings have been
460 // reduced to the underlying sequential or external string).
461 Label external_string, check_encoding;
462 __ Bind(&check_sequential);
463 STATIC_ASSERT(kSeqStringTag == 0);
464 __ TestAndBranchIfAnySet(result, kStringRepresentationMask, &external_string);
465
466 // Prepare sequential strings
467 STATIC_ASSERT(SeqTwoByteString::kHeaderSize == SeqOneByteString::kHeaderSize);
468 __ Add(string, string, SeqTwoByteString::kHeaderSize - kHeapObjectTag);
469 __ B(&check_encoding);
470
471 // Handle external strings.
472 __ Bind(&external_string);
473 if (FLAG_debug_code) {
474 // Assert that we do not have a cons or slice (indirect strings) here.
475 // Sequential strings have already been ruled out.
476 __ Tst(result, kIsIndirectStringMask);
477 __ Assert(eq, "external string expected, but not found");
478 }
479 // Rule out short external strings.
480 STATIC_CHECK(kShortExternalStringTag != 0);
481 __ TestAndBranchIfAnySet(result, kShortExternalStringMask, call_runtime);
482 __ Ldr(string, FieldMemOperand(string, ExternalString::kResourceDataOffset));
483
484 Label ascii, done;
485 __ Bind(&check_encoding);
486 STATIC_ASSERT(kTwoByteStringTag == 0);
487 __ TestAndBranchIfAnySet(result, kStringEncodingMask, &ascii);
488 // Two-byte string.
489 __ Ldrh(result, MemOperand(string, index, LSL, 1));
490 __ B(&done);
491 __ Bind(&ascii);
492 // Ascii string.
493 __ Ldrb(result, MemOperand(string, index));
494 __ Bind(&done);
495 }
496
497
498 static MemOperand ExpConstant(Register base, int index) {
499 return MemOperand(base, index * kDoubleSize);
500 }
501
502
503 void MathExpGenerator::EmitMathExp(MacroAssembler* masm,
504 DoubleRegister input,
505 DoubleRegister result,
506 DoubleRegister double_temp1,
507 DoubleRegister double_temp2,
508 Register temp1,
509 Register temp2,
510 Register temp3) {
511 // TODO(jbramley): There are several instances where fnmsub could be used
512 // instead of fmul and fsub. Doing this changes the result, but since this is
513 // an estimation anyway, does it matter?
514
515 ASSERT(!AreAliased(input, result,
516 double_temp1, double_temp2,
517 temp1, temp2, temp3));
518 ASSERT(ExternalReference::math_exp_constants(0).address() != NULL);
519
520 Label done;
521 DoubleRegister double_temp3 = result;
522 Register constants = temp3;
523
524 // The algorithm used relies on some magic constants which are initialized in
525 // ExternalReference::InitializeMathExpData().
526
527 // Load the address of the start of the array.
528 __ Mov(constants, Operand(ExternalReference::math_exp_constants(0)));
529
530 // We have to do a four-way split here:
531 // - If input <= about -708.4, the output always rounds to zero.
532 // - If input >= about 709.8, the output always rounds to +infinity.
533 // - If the input is NaN, the output is NaN.
534 // - Otherwise, the result needs to be calculated.
535 Label result_is_finite_non_zero;
536 // Assert that we can load offset 0 (the small input threshold) and offset 1
537 // (the large input threshold) with a single ldp.
538 ASSERT(kDRegSizeInBytes == (ExpConstant(constants, 1).offset() -
539 ExpConstant(constants, 0).offset()));
540 __ Ldp(double_temp1, double_temp2, ExpConstant(constants, 0));
541
542 __ Fcmp(input, double_temp1);
543 __ Fccmp(input, double_temp2, NoFlag, hi);
544 // At this point, the condition flags can be in one of five states:
545 // NZCV
546 // 1000 -708.4 < input < 709.8 result = exp(input)
547 // 0110 input == 709.8 result = +infinity
548 // 0010 input > 709.8 result = +infinity
549 // 0011 input is NaN result = input
550 // 0000 input <= -708.4 result = +0.0
551
552 // Continue the common case first. 'mi' tests N == 1.
553 __ B(&result_is_finite_non_zero, mi);
554
555 // TODO(jbramley): Add (and use) a zero D register for A64.
556 // TODO(jbramley): Consider adding a +infinity register for A64.
557 __ Ldr(double_temp2, ExpConstant(constants, 2)); // Synthesize +infinity.
558 __ Fsub(double_temp1, double_temp1, double_temp1); // Synthesize +0.0.
559
560 // Select between +0.0 and +infinity. 'lo' tests C == 0.
561 __ Fcsel(result, double_temp1, double_temp2, lo);
562 // Select between {+0.0 or +infinity} and input. 'vc' tests V == 0.
563 __ Fcsel(result, result, input, vc);
564 __ B(&done);
565
566 // The rest is magic, as described in InitializeMathExpData().
567 __ Bind(&result_is_finite_non_zero);
568
569 // Assert that we can load offset 3 and offset 4 with a single ldp.
570 ASSERT(kDRegSizeInBytes == (ExpConstant(constants, 4).offset() -
571 ExpConstant(constants, 3).offset()));
572 __ Ldp(double_temp1, double_temp3, ExpConstant(constants, 3));
573 __ Fmadd(double_temp1, double_temp1, input, double_temp3);
574 __ Fmov(temp2.W(), double_temp1.S());
575 __ Fsub(double_temp1, double_temp1, double_temp3);
576
577 // Assert that we can load offset 5 and offset 6 with a single ldp.
578 ASSERT(kDRegSizeInBytes == (ExpConstant(constants, 6).offset() -
579 ExpConstant(constants, 5).offset()));
580 __ Ldp(double_temp2, double_temp3, ExpConstant(constants, 5));
581 // TODO(jbramley): Consider using Fnmsub here.
582 __ Fmul(double_temp1, double_temp1, double_temp2);
583 __ Fsub(double_temp1, double_temp1, input);
584
585 __ Fmul(double_temp2, double_temp1, double_temp1);
586 __ Fsub(double_temp3, double_temp3, double_temp1);
587 __ Fmul(double_temp3, double_temp3, double_temp2);
588
589 __ Mov(temp1.W(), Operand(temp2.W(), LSR, 11));
590
591 __ Ldr(double_temp2, ExpConstant(constants, 7));
592 // TODO(jbramley): Consider using Fnmsub here.
593 __ Fmul(double_temp3, double_temp3, double_temp2);
594 __ Fsub(double_temp3, double_temp3, double_temp1);
595
596 // The 8th constant is 1.0, so use an immediate move rather than a load.
597 // We can't generate a runtime assertion here as we would need to call Abort
598 // in the runtime and we don't have an Isolate when we generate this code.
599 __ Fmov(double_temp2, 1.0);
600 __ Fadd(double_temp3, double_temp3, double_temp2);
601
602 __ And(temp2, temp2, 0x7ff);
603 __ Add(temp1, temp1, 0x3ff);
604
605 // Do the final table lookup.
606 __ Mov(temp3, Operand(ExternalReference::math_exp_log_table()));
607
608 __ Add(temp3, temp3, Operand(temp2, LSL, kDRegSizeInBytesLog2));
609 __ Ldp(temp2.W(), temp3.W(), MemOperand(temp3));
610 __ Orr(temp1.W(), temp3.W(), Operand(temp1.W(), LSL, 20));
611 __ Bfi(temp2, temp1, 32, 32);
612 __ Fmov(double_temp1, temp2);
613
614 __ Fmul(result, double_temp3, double_temp1);
615
616 __ Bind(&done);
617 }
618
619 #undef __
620
621 } } // namespace v8::internal
622
623 #endif // V8_TARGET_ARCH_A64
OLDNEW
« no previous file with comments | « src/a64/codegen-a64.h ('k') | src/a64/constants-a64.h » ('j') | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698