summaryrefslogtreecommitdiff
path: root/compiler/optimizing
diff options
context:
space:
mode:
Diffstat (limited to 'compiler/optimizing')
-rw-r--r--compiler/optimizing/code_generator_x86.cc24
-rw-r--r--compiler/optimizing/code_generator_x86_64.cc22
-rw-r--r--compiler/optimizing/intrinsics_x86.cc201
-rw-r--r--compiler/optimizing/intrinsics_x86_64.cc189
4 files changed, 385 insertions, 51 deletions
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index dc5c86efc6..9713d6a9c9 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -1312,7 +1312,7 @@ void InstructionCodeGeneratorX86::VisitCondition(HCondition* cond) {
}
// Convert the jumps into the result.
- Label done_label;
+ NearLabel done_label;
// False case: result = 0.
__ Bind(&false_label);
@@ -1968,7 +1968,7 @@ void InstructionCodeGeneratorX86::VisitTypeConversion(HTypeConversion* conversio
XmmRegister input = in.AsFpuRegister<XmmRegister>();
Register output = out.AsRegister<Register>();
XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
- Label done, nan;
+ NearLabel done, nan;
__ movl(output, Immediate(kPrimIntMax));
// temp = int-to-float(output)
@@ -1993,7 +1993,7 @@ void InstructionCodeGeneratorX86::VisitTypeConversion(HTypeConversion* conversio
XmmRegister input = in.AsFpuRegister<XmmRegister>();
Register output = out.AsRegister<Register>();
XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
- Label done, nan;
+ NearLabel done, nan;
__ movl(output, Immediate(kPrimIntMax));
// temp = int-to-double(output)
@@ -2652,7 +2652,7 @@ void InstructionCodeGeneratorX86::GenerateRemFP(HRem *rem) {
PushOntoFPStack(first, 0, 2 * elem_size, /* is_fp */ true, is_wide);
// Loop doing FPREM until we stabilize.
- Label retry;
+ NearLabel retry;
__ Bind(&retry);
__ fprem();
@@ -2766,8 +2766,8 @@ void InstructionCodeGeneratorX86::GenerateDivRemWithAnyConstant(HBinaryOperation
int shift;
CalculateMagicAndShiftForDivRem(imm, false /* is_long */, &magic, &shift);
- Label ndiv;
- Label end;
+ NearLabel ndiv;
+ NearLabel end;
// If numerator is 0, the result is 0, no computation needed.
__ testl(eax, eax);
__ j(kNotEqual, &ndiv);
@@ -3243,7 +3243,7 @@ void InstructionCodeGeneratorX86::GenerateShlLong(const Location& loc, int shift
}
void InstructionCodeGeneratorX86::GenerateShlLong(const Location& loc, Register shifter) {
- Label done;
+ NearLabel done;
__ shld(loc.AsRegisterPairHigh<Register>(), loc.AsRegisterPairLow<Register>(), shifter);
__ shll(loc.AsRegisterPairLow<Register>(), shifter);
__ testl(shifter, Immediate(32));
@@ -3275,7 +3275,7 @@ void InstructionCodeGeneratorX86::GenerateShrLong(const Location& loc, int shift
}
void InstructionCodeGeneratorX86::GenerateShrLong(const Location& loc, Register shifter) {
- Label done;
+ NearLabel done;
__ shrd(loc.AsRegisterPairLow<Register>(), loc.AsRegisterPairHigh<Register>(), shifter);
__ sarl(loc.AsRegisterPairHigh<Register>(), shifter);
__ testl(shifter, Immediate(32));
@@ -3310,7 +3310,7 @@ void InstructionCodeGeneratorX86::GenerateUShrLong(const Location& loc, int shif
}
void InstructionCodeGeneratorX86::GenerateUShrLong(const Location& loc, Register shifter) {
- Label done;
+ NearLabel done;
__ shrd(loc.AsRegisterPairLow<Register>(), loc.AsRegisterPairHigh<Register>(), shifter);
__ shrl(loc.AsRegisterPairHigh<Register>(), shifter);
__ testl(shifter, Immediate(32));
@@ -3485,7 +3485,7 @@ void InstructionCodeGeneratorX86::VisitCompare(HCompare* compare) {
Location left = locations->InAt(0);
Location right = locations->InAt(1);
- Label less, greater, done;
+ NearLabel less, greater, done;
switch (compare->InputAt(0)->GetType()) {
case Primitive::kPrimLong: {
Register left_low = left.AsRegisterPairLow<Register>();
@@ -3709,7 +3709,7 @@ void CodeGeneratorX86::MarkGCCard(Register temp,
Register object,
Register value,
bool value_can_be_null) {
- Label is_null;
+ NearLabel is_null;
if (value_can_be_null) {
__ testl(value, value);
__ j(kEqual, &is_null);
@@ -4946,7 +4946,7 @@ void InstructionCodeGeneratorX86::VisitInstanceOf(HInstanceOf* instruction) {
Location cls = locations->InAt(1);
Register out = locations->Out().AsRegister<Register>();
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
- Label done, zero;
+ NearLabel done, zero;
SlowPathCodeX86* slow_path = nullptr;
// Return 0 if `obj` is null.
diff --git a/compiler/optimizing/code_generator_x86_64.cc b/compiler/optimizing/code_generator_x86_64.cc
index 0cf1089cf8..43a3e52a7f 100644
--- a/compiler/optimizing/code_generator_x86_64.cc
+++ b/compiler/optimizing/code_generator_x86_64.cc
@@ -1324,7 +1324,7 @@ void InstructionCodeGeneratorX86_64::VisitCondition(HCondition* cond) {
}
// Convert the jumps into the result.
- Label done_label;
+ NearLabel done_label;
// False case: result = 0.
__ Bind(&false_label);
@@ -1413,7 +1413,7 @@ void InstructionCodeGeneratorX86_64::VisitCompare(HCompare* compare) {
Location left = locations->InAt(0);
Location right = locations->InAt(1);
- Label less, greater, done;
+ NearLabel less, greater, done;
Primitive::Type type = compare->InputAt(0)->GetType();
switch (type) {
case Primitive::kPrimLong: {
@@ -2123,7 +2123,7 @@ void InstructionCodeGeneratorX86_64::VisitTypeConversion(HTypeConversion* conver
// Processing a Dex `float-to-int' instruction.
XmmRegister input = in.AsFpuRegister<XmmRegister>();
CpuRegister output = out.AsRegister<CpuRegister>();
- Label done, nan;
+ NearLabel done, nan;
__ movl(output, Immediate(kPrimIntMax));
// if input >= (float)INT_MAX goto done
@@ -2145,7 +2145,7 @@ void InstructionCodeGeneratorX86_64::VisitTypeConversion(HTypeConversion* conver
// Processing a Dex `double-to-int' instruction.
XmmRegister input = in.AsFpuRegister<XmmRegister>();
CpuRegister output = out.AsRegister<CpuRegister>();
- Label done, nan;
+ NearLabel done, nan;
__ movl(output, Immediate(kPrimIntMax));
// if input >= (double)INT_MAX goto done
@@ -2187,7 +2187,7 @@ void InstructionCodeGeneratorX86_64::VisitTypeConversion(HTypeConversion* conver
// Processing a Dex `float-to-long' instruction.
XmmRegister input = in.AsFpuRegister<XmmRegister>();
CpuRegister output = out.AsRegister<CpuRegister>();
- Label done, nan;
+ NearLabel done, nan;
codegen_->Load64BitValue(output, kPrimLongMax);
// if input >= (float)LONG_MAX goto done
@@ -2209,7 +2209,7 @@ void InstructionCodeGeneratorX86_64::VisitTypeConversion(HTypeConversion* conver
// Processing a Dex `double-to-long' instruction.
XmmRegister input = in.AsFpuRegister<XmmRegister>();
CpuRegister output = out.AsRegister<CpuRegister>();
- Label done, nan;
+ NearLabel done, nan;
codegen_->Load64BitValue(output, kPrimLongMax);
// if input >= (double)LONG_MAX goto done
@@ -2772,7 +2772,7 @@ void InstructionCodeGeneratorX86_64::GenerateRemFP(HRem *rem) {
PushOntoFPStack(first, 0, 2 * elem_size, is_float);
// Loop doing FPREM until we stabilize.
- Label retry;
+ NearLabel retry;
__ Bind(&retry);
__ fprem();
@@ -2926,8 +2926,8 @@ void InstructionCodeGeneratorX86_64::GenerateDivRemWithAnyConstant(HBinaryOperat
__ movl(numerator, eax);
- Label no_div;
- Label end;
+ NearLabel no_div;
+ NearLabel end;
__ testl(eax, eax);
__ j(kNotEqual, &no_div);
@@ -4247,7 +4247,7 @@ void CodeGeneratorX86_64::MarkGCCard(CpuRegister temp,
CpuRegister object,
CpuRegister value,
bool value_can_be_null) {
- Label is_null;
+ NearLabel is_null;
if (value_can_be_null) {
__ testl(value, value);
__ j(kEqual, &is_null);
@@ -4674,7 +4674,7 @@ void InstructionCodeGeneratorX86_64::VisitInstanceOf(HInstanceOf* instruction) {
Location cls = locations->InAt(1);
CpuRegister out = locations->Out().AsRegister<CpuRegister>();
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
- Label done, zero;
+ NearLabel done, zero;
SlowPathCodeX86_64* slow_path = nullptr;
// Return 0 if `obj` is null.
diff --git a/compiler/optimizing/intrinsics_x86.cc b/compiler/optimizing/intrinsics_x86.cc
index c5d88d2b25..e302317d14 100644
--- a/compiler/optimizing/intrinsics_x86.cc
+++ b/compiler/optimizing/intrinsics_x86.cc
@@ -507,7 +507,7 @@ static void GenMinMaxFP(LocationSummary* locations, bool is_min, bool is_double,
XmmRegister op2 = op2_loc.AsFpuRegister<XmmRegister>();
- Label nan, done, op2_label;
+ NearLabel nan, done, op2_label;
if (is_double) {
__ ucomisd(out, op2);
} else {
@@ -841,7 +841,7 @@ void IntrinsicCodeGeneratorX86::VisitMathRoundFloat(HInvoke* invoke) {
Register out = locations->Out().AsRegister<Register>();
XmmRegister maxInt = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
XmmRegister inPlusPointFive = locations->GetTemp(1).AsFpuRegister<XmmRegister>();
- Label done, nan;
+ NearLabel done, nan;
X86Assembler* assembler = GetAssembler();
// Generate 0.5 into inPlusPointFive.
@@ -888,9 +888,9 @@ void IntrinsicLocationsBuilderX86::VisitStringCharAt(HInvoke* invoke) {
void IntrinsicCodeGeneratorX86::VisitStringCharAt(HInvoke* invoke) {
LocationSummary* locations = invoke->GetLocations();
- // Location of reference to data array
+ // Location of reference to data array.
const int32_t value_offset = mirror::String::ValueOffset().Int32Value();
- // Location of count
+ // Location of count.
const int32_t count_offset = mirror::String::CountOffset().Int32Value();
Register obj = locations->InAt(0).AsRegister<Register>();
@@ -917,6 +917,183 @@ void IntrinsicCodeGeneratorX86::VisitStringCharAt(HInvoke* invoke) {
__ Bind(slow_path->GetExitLabel());
}
+void IntrinsicLocationsBuilderX86::VisitSystemArrayCopyChar(HInvoke* invoke) {
+ // We need at least two of the positions or length to be an integer constant,
+ // or else we won't have enough free registers.
+ HIntConstant* src_pos = invoke->InputAt(1)->AsIntConstant();
+ HIntConstant* dest_pos = invoke->InputAt(3)->AsIntConstant();
+ HIntConstant* length = invoke->InputAt(4)->AsIntConstant();
+
+ int num_constants =
+ ((src_pos != nullptr) ? 1 : 0)
+ + ((dest_pos != nullptr) ? 1 : 0)
+ + ((length != nullptr) ? 1 : 0);
+
+ if (num_constants < 2) {
+ // Not enough free registers.
+ return;
+ }
+
+ // As long as we are checking, we might as well check to see if the src and dest
+ // positions are >= 0.
+ if ((src_pos != nullptr && src_pos->GetValue() < 0) ||
+ (dest_pos != nullptr && dest_pos->GetValue() < 0)) {
+ // We will have to fail anyways.
+ return;
+ }
+
+ // And since we are already checking, check the length too.
+ if (length != nullptr) {
+ int32_t len = length->GetValue();
+ if (len < 0) {
+ // Just call as normal.
+ return;
+ }
+ }
+
+ // Okay, it is safe to generate inline code.
+ LocationSummary* locations =
+ new (arena_) LocationSummary(invoke, LocationSummary::kCallOnSlowPath, kIntrinsified);
+ // arraycopy(Object src, int srcPos, Object dest, int destPos, int length).
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetInAt(1, Location::RegisterOrConstant(invoke->InputAt(1)));
+ locations->SetInAt(2, Location::RequiresRegister());
+ locations->SetInAt(3, Location::RegisterOrConstant(invoke->InputAt(3)));
+ locations->SetInAt(4, Location::RegisterOrConstant(invoke->InputAt(4)));
+
+ // And we need some temporaries. We will use REP MOVSW, so we need fixed registers.
+ locations->AddTemp(Location::RegisterLocation(ESI));
+ locations->AddTemp(Location::RegisterLocation(EDI));
+ locations->AddTemp(Location::RegisterLocation(ECX));
+}
+
+static void CheckPosition(X86Assembler* assembler,
+ Location pos,
+ Register input,
+ Register length,
+ SlowPathCodeX86* slow_path,
+ Register input_len,
+ Register temp) {
+ // Where is the length in the String?
+ const uint32_t length_offset = mirror::Array::LengthOffset().Uint32Value();
+
+ if (pos.IsConstant()) {
+ int32_t pos_const = pos.GetConstant()->AsIntConstant()->GetValue();
+ if (pos_const == 0) {
+ // Check that length(input) >= length.
+ __ cmpl(Address(input, length_offset), length);
+ __ j(kLess, slow_path->GetEntryLabel());
+ } else {
+ // Check that length(input) >= pos.
+ __ movl(input_len, Address(input, length_offset));
+ __ cmpl(input_len, Immediate(pos_const));
+ __ j(kLess, slow_path->GetEntryLabel());
+
+ // Check that (length(input) - pos) >= length.
+ __ leal(temp, Address(input_len, -pos_const));
+ __ cmpl(temp, length);
+ __ j(kLess, slow_path->GetEntryLabel());
+ }
+ } else {
+ // Check that pos >= 0.
+ Register pos_reg = pos.AsRegister<Register>();
+ __ testl(pos_reg, pos_reg);
+ __ j(kLess, slow_path->GetEntryLabel());
+
+ // Check that pos <= length(input).
+ __ cmpl(Address(input, length_offset), pos_reg);
+ __ j(kLess, slow_path->GetEntryLabel());
+
+ // Check that (length(input) - pos) >= length.
+ __ movl(temp, Address(input, length_offset));
+ __ subl(temp, pos_reg);
+ __ cmpl(temp, length);
+ __ j(kLess, slow_path->GetEntryLabel());
+ }
+}
+
+void IntrinsicCodeGeneratorX86::VisitSystemArrayCopyChar(HInvoke* invoke) {
+ X86Assembler* assembler = GetAssembler();
+ LocationSummary* locations = invoke->GetLocations();
+
+ Register src = locations->InAt(0).AsRegister<Register>();
+ Location srcPos = locations->InAt(1);
+ Register dest = locations->InAt(2).AsRegister<Register>();
+ Location destPos = locations->InAt(3);
+ Location length = locations->InAt(4);
+
+ // Temporaries that we need for MOVSW.
+ Register src_base = locations->GetTemp(0).AsRegister<Register>();
+ DCHECK_EQ(src_base, ESI);
+ Register dest_base = locations->GetTemp(1).AsRegister<Register>();
+ DCHECK_EQ(dest_base, EDI);
+ Register count = locations->GetTemp(2).AsRegister<Register>();
+ DCHECK_EQ(count, ECX);
+
+ SlowPathCodeX86* slow_path = new (GetAllocator()) IntrinsicSlowPathX86(invoke);
+ codegen_->AddSlowPath(slow_path);
+
+ // Bail out if the source and destination are the same (to handle overlap).
+ __ cmpl(src, dest);
+ __ j(kEqual, slow_path->GetEntryLabel());
+
+ // Bail out if the source is null.
+ __ testl(src, src);
+ __ j(kEqual, slow_path->GetEntryLabel());
+
+ // Bail out if the destination is null.
+ __ testl(dest, dest);
+ __ j(kEqual, slow_path->GetEntryLabel());
+
+ // If the length is negative, bail out.
+ // We have already checked in the LocationsBuilder for the constant case.
+ if (!length.IsConstant()) {
+ __ cmpl(length.AsRegister<Register>(), length.AsRegister<Register>());
+ __ j(kLess, slow_path->GetEntryLabel());
+ }
+
+ // We need the count in ECX.
+ if (length.IsConstant()) {
+ __ movl(count, Immediate(length.GetConstant()->AsIntConstant()->GetValue()));
+ } else {
+ __ movl(count, length.AsRegister<Register>());
+ }
+
+ // Validity checks: source.
+ CheckPosition(assembler, srcPos, src, count, slow_path, src_base, dest_base);
+
+ // Validity checks: dest.
+ CheckPosition(assembler, destPos, dest, count, slow_path, src_base, dest_base);
+
+ // Okay, everything checks out. Finally time to do the copy.
+ // Check assumption that sizeof(Char) is 2 (used in scaling below).
+ const size_t char_size = Primitive::ComponentSize(Primitive::kPrimChar);
+ DCHECK_EQ(char_size, 2u);
+
+ const uint32_t data_offset = mirror::Array::DataOffset(char_size).Uint32Value();
+
+ if (srcPos.IsConstant()) {
+ int32_t srcPos_const = srcPos.GetConstant()->AsIntConstant()->GetValue();
+ __ leal(src_base, Address(src, char_size * srcPos_const + data_offset));
+ } else {
+ __ leal(src_base, Address(src, srcPos.AsRegister<Register>(),
+ ScaleFactor::TIMES_2, data_offset));
+ }
+ if (destPos.IsConstant()) {
+ int32_t destPos_const = destPos.GetConstant()->AsIntConstant()->GetValue();
+
+ __ leal(dest_base, Address(dest, char_size * destPos_const + data_offset));
+ } else {
+ __ leal(dest_base, Address(dest, destPos.AsRegister<Register>(),
+ ScaleFactor::TIMES_2, data_offset));
+ }
+
+ // Do the move.
+ __ rep_movsw();
+
+ __ Bind(slow_path->GetExitLabel());
+}
+
void IntrinsicLocationsBuilderX86::VisitStringCompareTo(HInvoke* invoke) {
// The inputs plus one temp.
LocationSummary* locations = new (arena_) LocationSummary(invoke,
@@ -970,9 +1147,7 @@ void IntrinsicCodeGeneratorX86::VisitStringEquals(HInvoke* invoke) {
Register edi = locations->GetTemp(1).AsRegister<Register>();
Register esi = locations->Out().AsRegister<Register>();
- Label end;
- Label return_true;
- Label return_false;
+ NearLabel end, return_true, return_false;
// Get offsets of count, value, and class fields within a string object.
const uint32_t count_offset = mirror::String::CountOffset().Uint32Value();
@@ -1004,8 +1179,7 @@ void IntrinsicCodeGeneratorX86::VisitStringEquals(HInvoke* invoke) {
__ cmpl(ecx, Address(arg, count_offset));
__ j(kNotEqual, &return_false);
// Return true if both strings are empty.
- __ testl(ecx, ecx);
- __ j(kEqual, &return_true);
+ __ jecxz(&return_true);
// Load starting addresses of string values into ESI/EDI as required for repe_cmpsl instruction.
__ leal(esi, Address(str, value_offset));
@@ -1115,7 +1289,7 @@ static void GenerateStringIndexOf(HInvoke* invoke,
// Do a zero-length check.
// TODO: Support jecxz.
- Label not_found_label;
+ NearLabel not_found_label;
__ testl(string_length, string_length);
__ j(kEqual, &not_found_label);
@@ -1158,7 +1332,7 @@ static void GenerateStringIndexOf(HInvoke* invoke,
__ subl(string_length, counter);
__ leal(out, Address(string_length, -1));
- Label done;
+ NearLabel done;
__ jmp(&done);
// Failed to match; return -1.
@@ -1878,7 +2052,7 @@ static void GenLeadingZeros(X86Assembler* assembler, HInvoke* invoke, bool is_lo
}
// BSR sets ZF if the input was zero, and the output is undefined.
- Label all_zeroes, done;
+ NearLabel all_zeroes, done;
__ j(kEqual, &all_zeroes);
// Correct the result from BSR to get the final CLZ result.
@@ -1897,7 +2071,7 @@ static void GenLeadingZeros(X86Assembler* assembler, HInvoke* invoke, bool is_lo
DCHECK(src.IsRegisterPair());
Register src_lo = src.AsRegisterPairLow<Register>();
Register src_hi = src.AsRegisterPairHigh<Register>();
- Label handle_low, done, all_zeroes;
+ NearLabel handle_low, done, all_zeroes;
// Is the high word zero?
__ testl(src_hi, src_hi);
@@ -1954,7 +2128,6 @@ void IntrinsicCodeGeneratorX86::Visit ## Name(HInvoke* invoke ATTRIBUTE_UNUSED)
UNIMPLEMENTED_INTRINSIC(MathRoundDouble)
UNIMPLEMENTED_INTRINSIC(StringGetCharsNoCheck)
-UNIMPLEMENTED_INTRINSIC(SystemArrayCopyChar)
UNIMPLEMENTED_INTRINSIC(ReferenceGetReferent)
UNIMPLEMENTED_INTRINSIC(IntegerNumberOfTrailingZeros)
UNIMPLEMENTED_INTRINSIC(LongNumberOfTrailingZeros)
diff --git a/compiler/optimizing/intrinsics_x86_64.cc b/compiler/optimizing/intrinsics_x86_64.cc
index 258ae9a55f..51980af36d 100644
--- a/compiler/optimizing/intrinsics_x86_64.cc
+++ b/compiler/optimizing/intrinsics_x86_64.cc
@@ -405,7 +405,7 @@ static void GenMinMaxFP(LocationSummary* locations,
XmmRegister op2 = op2_loc.AsFpuRegister<XmmRegister>();
- Label nan, done, op2_label;
+ NearLabel nan, done, op2_label;
if (is_double) {
__ ucomisd(out, op2);
} else {
@@ -702,7 +702,7 @@ void IntrinsicCodeGeneratorX86_64::VisitMathRoundFloat(HInvoke* invoke) {
XmmRegister in = locations->InAt(0).AsFpuRegister<XmmRegister>();
CpuRegister out = locations->Out().AsRegister<CpuRegister>();
XmmRegister inPlusPointFive = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
- Label done, nan;
+ NearLabel done, nan;
X86_64Assembler* assembler = GetAssembler();
// Load 0.5 into inPlusPointFive.
@@ -750,7 +750,7 @@ void IntrinsicCodeGeneratorX86_64::VisitMathRoundDouble(HInvoke* invoke) {
XmmRegister in = locations->InAt(0).AsFpuRegister<XmmRegister>();
CpuRegister out = locations->Out().AsRegister<CpuRegister>();
XmmRegister inPlusPointFive = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
- Label done, nan;
+ NearLabel done, nan;
X86_64Assembler* assembler = GetAssembler();
// Load 0.5 into inPlusPointFive.
@@ -797,9 +797,9 @@ void IntrinsicLocationsBuilderX86_64::VisitStringCharAt(HInvoke* invoke) {
void IntrinsicCodeGeneratorX86_64::VisitStringCharAt(HInvoke* invoke) {
LocationSummary* locations = invoke->GetLocations();
- // Location of reference to data array
+ // Location of reference to data array.
const int32_t value_offset = mirror::String::ValueOffset().Int32Value();
- // Location of count
+ // Location of count.
const int32_t count_offset = mirror::String::CountOffset().Int32Value();
CpuRegister obj = locations->InAt(0).AsRegister<CpuRegister>();
@@ -826,6 +826,171 @@ void IntrinsicCodeGeneratorX86_64::VisitStringCharAt(HInvoke* invoke) {
__ Bind(slow_path->GetExitLabel());
}
+void IntrinsicLocationsBuilderX86_64::VisitSystemArrayCopyChar(HInvoke* invoke) {
+ // Check to see if we have known failures that will cause us to have to bail out
+ // to the runtime, and just generate the runtime call directly.
+ HIntConstant* src_pos = invoke->InputAt(1)->AsIntConstant();
+ HIntConstant* dest_pos = invoke->InputAt(3)->AsIntConstant();
+
+ // The positions must be non-negative.
+ if ((src_pos != nullptr && src_pos->GetValue() < 0) ||
+ (dest_pos != nullptr && dest_pos->GetValue() < 0)) {
+ // We will have to fail anyways.
+ return;
+ }
+
+ // The length must be > 0.
+ HIntConstant* length = invoke->InputAt(4)->AsIntConstant();
+ if (length != nullptr) {
+ int32_t len = length->GetValue();
+ if (len < 0) {
+ // Just call as normal.
+ return;
+ }
+ }
+
+ LocationSummary* locations = new (arena_) LocationSummary(invoke,
+ LocationSummary::kCallOnSlowPath,
+ kIntrinsified);
+ // arraycopy(Object src, int srcPos, Object dest, int destPos, int length).
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetInAt(1, Location::RegisterOrConstant(invoke->InputAt(1)));
+ locations->SetInAt(2, Location::RequiresRegister());
+ locations->SetInAt(3, Location::RegisterOrConstant(invoke->InputAt(3)));
+ locations->SetInAt(4, Location::RegisterOrConstant(invoke->InputAt(4)));
+
+ // And we need some temporaries. We will use REP MOVSW, so we need fixed registers.
+ locations->AddTemp(Location::RegisterLocation(RSI));
+ locations->AddTemp(Location::RegisterLocation(RDI));
+ locations->AddTemp(Location::RegisterLocation(RCX));
+}
+
+static void CheckPosition(X86_64Assembler* assembler,
+ Location pos,
+ CpuRegister input,
+ CpuRegister length,
+ SlowPathCodeX86_64* slow_path,
+ CpuRegister input_len,
+ CpuRegister temp) {
+ // Where is the length in the String?
+ const uint32_t length_offset = mirror::Array::LengthOffset().Uint32Value();
+
+ if (pos.IsConstant()) {
+ int32_t pos_const = pos.GetConstant()->AsIntConstant()->GetValue();
+ if (pos_const == 0) {
+ // Check that length(input) >= length.
+ __ cmpl(Address(input, length_offset), length);
+ __ j(kLess, slow_path->GetEntryLabel());
+ } else {
+ // Check that length(input) >= pos.
+ __ movl(input_len, Address(input, length_offset));
+ __ cmpl(input_len, Immediate(pos_const));
+ __ j(kLess, slow_path->GetEntryLabel());
+
+ // Check that (length(input) - pos) >= length.
+ __ leal(temp, Address(input_len, -pos_const));
+ __ cmpl(temp, length);
+ __ j(kLess, slow_path->GetEntryLabel());
+ }
+ } else {
+ // Check that pos >= 0.
+ CpuRegister pos_reg = pos.AsRegister<CpuRegister>();
+ __ testl(pos_reg, pos_reg);
+ __ j(kLess, slow_path->GetEntryLabel());
+
+ // Check that pos <= length(input).
+ __ cmpl(Address(input, length_offset), pos_reg);
+ __ j(kLess, slow_path->GetEntryLabel());
+
+ // Check that (length(input) - pos) >= length.
+ __ movl(temp, Address(input, length_offset));
+ __ subl(temp, pos_reg);
+ __ cmpl(temp, length);
+ __ j(kLess, slow_path->GetEntryLabel());
+ }
+}
+
+void IntrinsicCodeGeneratorX86_64::VisitSystemArrayCopyChar(HInvoke* invoke) {
+ X86_64Assembler* assembler = GetAssembler();
+ LocationSummary* locations = invoke->GetLocations();
+
+ CpuRegister src = locations->InAt(0).AsRegister<CpuRegister>();
+ Location srcPos = locations->InAt(1);
+ CpuRegister dest = locations->InAt(2).AsRegister<CpuRegister>();
+ Location destPos = locations->InAt(3);
+ Location length = locations->InAt(4);
+
+ // Temporaries that we need for MOVSW.
+ CpuRegister src_base = locations->GetTemp(0).AsRegister<CpuRegister>();
+ DCHECK_EQ(src_base.AsRegister(), RSI);
+ CpuRegister dest_base = locations->GetTemp(1).AsRegister<CpuRegister>();
+ DCHECK_EQ(dest_base.AsRegister(), RDI);
+ CpuRegister count = locations->GetTemp(2).AsRegister<CpuRegister>();
+ DCHECK_EQ(count.AsRegister(), RCX);
+
+ SlowPathCodeX86_64* slow_path = new (GetAllocator()) IntrinsicSlowPathX86_64(invoke);
+ codegen_->AddSlowPath(slow_path);
+
+ // Bail out if the source and destination are the same.
+ __ cmpl(src, dest);
+ __ j(kEqual, slow_path->GetEntryLabel());
+
+ // Bail out if the source is null.
+ __ testl(src, src);
+ __ j(kEqual, slow_path->GetEntryLabel());
+
+ // Bail out if the destination is null.
+ __ testl(dest, dest);
+ __ j(kEqual, slow_path->GetEntryLabel());
+
+ // If the length is negative, bail out.
+ // We have already checked in the LocationsBuilder for the constant case.
+ if (!length.IsConstant()) {
+ __ testl(length.AsRegister<CpuRegister>(), length.AsRegister<CpuRegister>());
+ __ j(kLess, slow_path->GetEntryLabel());
+ }
+
+ // We need the count in RCX.
+ if (length.IsConstant()) {
+ __ movl(count, Immediate(length.GetConstant()->AsIntConstant()->GetValue()));
+ } else {
+ __ movl(count, length.AsRegister<CpuRegister>());
+ }
+
+ // Validity checks: source.
+ CheckPosition(assembler, srcPos, src, count, slow_path, src_base, dest_base);
+
+ // Validity checks: dest.
+ CheckPosition(assembler, destPos, dest, count, slow_path, src_base, dest_base);
+
+ // Okay, everything checks out. Finally time to do the copy.
+ // Check assumption that sizeof(Char) is 2 (used in scaling below).
+ const size_t char_size = Primitive::ComponentSize(Primitive::kPrimChar);
+ DCHECK_EQ(char_size, 2u);
+
+ const uint32_t data_offset = mirror::Array::DataOffset(char_size).Uint32Value();
+
+ if (srcPos.IsConstant()) {
+ int32_t srcPos_const = srcPos.GetConstant()->AsIntConstant()->GetValue();
+ __ leal(src_base, Address(src, char_size * srcPos_const + data_offset));
+ } else {
+ __ leal(src_base, Address(src, srcPos.AsRegister<CpuRegister>(),
+ ScaleFactor::TIMES_2, data_offset));
+ }
+ if (destPos.IsConstant()) {
+ int32_t destPos_const = destPos.GetConstant()->AsIntConstant()->GetValue();
+ __ leal(dest_base, Address(dest, char_size * destPos_const + data_offset));
+ } else {
+ __ leal(dest_base, Address(dest, destPos.AsRegister<CpuRegister>(),
+ ScaleFactor::TIMES_2, data_offset));
+ }
+
+ // Do the move.
+ __ rep_movsw();
+
+ __ Bind(slow_path->GetExitLabel());
+}
+
void IntrinsicLocationsBuilderX86_64::VisitStringCompareTo(HInvoke* invoke) {
LocationSummary* locations = new (arena_) LocationSummary(invoke,
LocationSummary::kCall,
@@ -879,9 +1044,7 @@ void IntrinsicCodeGeneratorX86_64::VisitStringEquals(HInvoke* invoke) {
CpuRegister rdi = locations->GetTemp(1).AsRegister<CpuRegister>();
CpuRegister rsi = locations->Out().AsRegister<CpuRegister>();
- Label end;
- Label return_true;
- Label return_false;
+ NearLabel end, return_true, return_false;
// Get offsets of count, value, and class fields within a string object.
const uint32_t count_offset = mirror::String::CountOffset().Uint32Value();
@@ -913,8 +1076,7 @@ void IntrinsicCodeGeneratorX86_64::VisitStringEquals(HInvoke* invoke) {
__ cmpl(rcx, Address(arg, count_offset));
__ j(kNotEqual, &return_false);
// Return true if both strings are empty.
- __ testl(rcx, rcx);
- __ j(kEqual, &return_true);
+ __ jrcxz(&return_true);
// Load starting addresses of string values into RSI/RDI as required for repe_cmpsq instruction.
__ leal(rsi, Address(str, value_offset));
@@ -1024,7 +1186,7 @@ static void GenerateStringIndexOf(HInvoke* invoke,
// Do a length check.
// TODO: Support jecxz.
- Label not_found_label;
+ NearLabel not_found_label;
__ testl(string_length, string_length);
__ j(kEqual, &not_found_label);
@@ -1066,7 +1228,7 @@ static void GenerateStringIndexOf(HInvoke* invoke,
__ subl(string_length, counter);
__ leal(out, Address(string_length, -1));
- Label done;
+ NearLabel done;
__ jmp(&done);
// Failed to match; return -1.
@@ -1731,7 +1893,7 @@ static void GenLeadingZeros(X86_64Assembler* assembler, HInvoke* invoke, bool is
}
// BSR sets ZF if the input was zero, and the output is undefined.
- Label is_zero, done;
+ NearLabel is_zero, done;
__ j(kEqual, &is_zero);
// Correct the result from BSR to get the CLZ result.
@@ -1772,7 +1934,6 @@ void IntrinsicCodeGeneratorX86_64::Visit ## Name(HInvoke* invoke ATTRIBUTE_UNUSE
}
UNIMPLEMENTED_INTRINSIC(StringGetCharsNoCheck)
-UNIMPLEMENTED_INTRINSIC(SystemArrayCopyChar)
UNIMPLEMENTED_INTRINSIC(ReferenceGetReferent)
UNIMPLEMENTED_INTRINSIC(IntegerNumberOfTrailingZeros)
UNIMPLEMENTED_INTRINSIC(LongNumberOfTrailingZeros)
generated by cgit v1.2.3-59-g8ed1b (git 2.39.0) at 2025-10-21 14:19:35 +0000