ARM: Specify if some branches go to far targets
Also avoids some branches to branches.
Test: m test-art-target
Change-Id: I839bcaa6ba5fba7aaa8474bee5532366ac0cc4ac
diff --git a/compiler/optimizing/code_generator_arm_vixl.cc b/compiler/optimizing/code_generator_arm_vixl.cc
index 34821f8..1f8e1ef 100644
--- a/compiler/optimizing/code_generator_arm_vixl.cc
+++ b/compiler/optimizing/code_generator_arm_vixl.cc
@@ -2139,7 +2139,8 @@
static void GenerateLongComparesAndJumps(HCondition* cond,
vixl32::Label* true_label,
vixl32::Label* false_label,
- CodeGeneratorARMVIXL* codegen) {
+ CodeGeneratorARMVIXL* codegen,
+ bool is_far_target = true) {
LocationSummary* locations = cond->GetLocations();
Location left = locations->InAt(0);
Location right = locations->InAt(1);
@@ -2190,12 +2191,12 @@
__ Cmp(left_high, val_high);
if (if_cond == kCondNE) {
- __ B(ARMCondition(true_high_cond), true_label);
+ __ B(ARMCondition(true_high_cond), true_label, is_far_target);
} else if (if_cond == kCondEQ) {
- __ B(ARMCondition(false_high_cond), false_label);
+ __ B(ARMCondition(false_high_cond), false_label, is_far_target);
} else {
- __ B(ARMCondition(true_high_cond), true_label);
- __ B(ARMCondition(false_high_cond), false_label);
+ __ B(ARMCondition(true_high_cond), true_label, is_far_target);
+ __ B(ARMCondition(false_high_cond), false_label, is_far_target);
}
// Must be equal high, so compare the lows.
__ Cmp(left_low, val_low);
@@ -2205,19 +2206,19 @@
__ Cmp(left_high, right_high);
if (if_cond == kCondNE) {
- __ B(ARMCondition(true_high_cond), true_label);
+ __ B(ARMCondition(true_high_cond), true_label, is_far_target);
} else if (if_cond == kCondEQ) {
- __ B(ARMCondition(false_high_cond), false_label);
+ __ B(ARMCondition(false_high_cond), false_label, is_far_target);
} else {
- __ B(ARMCondition(true_high_cond), true_label);
- __ B(ARMCondition(false_high_cond), false_label);
+ __ B(ARMCondition(true_high_cond), true_label, is_far_target);
+ __ B(ARMCondition(false_high_cond), false_label, is_far_target);
}
// Must be equal high, so compare the lows.
__ Cmp(left_low, right_low);
}
// The last comparison might be unsigned.
// TODO: optimize cases where this is always true/false
- __ B(final_condition, true_label);
+ __ B(final_condition, true_label, is_far_target);
}
static void GenerateConditionLong(HCondition* cond, CodeGeneratorARMVIXL* codegen) {
@@ -2292,7 +2293,7 @@
vixl32::Label* const final_label = codegen->GetFinalLabel(cond, &done_label);
vixl32::Label true_label, false_label;
- GenerateLongComparesAndJumps(cond, &true_label, &false_label, codegen);
+ GenerateLongComparesAndJumps(cond, &true_label, &false_label, codegen, /* is_far_target */ false);
// False case: result = 0.
__ Bind(&false_label);
@@ -2957,7 +2958,8 @@
void InstructionCodeGeneratorARMVIXL::GenerateCompareTestAndBranch(HCondition* condition,
vixl32::Label* true_target_in,
- vixl32::Label* false_target_in) {
+ vixl32::Label* false_target_in,
+ bool is_far_target) {
if (CanGenerateTest(condition, codegen_->GetAssembler())) {
vixl32::Label* non_fallthrough_target;
bool invert;
@@ -2973,7 +2975,7 @@
const auto cond = GenerateTest(condition, invert, codegen_);
- __ B(cond.first, non_fallthrough_target);
+ __ B(cond.first, non_fallthrough_target, is_far_target);
if (false_target_in != nullptr && false_target_in != non_fallthrough_target) {
__ B(false_target_in);
@@ -2989,7 +2991,7 @@
vixl32::Label* false_target = (false_target_in == nullptr) ? &fallthrough : false_target_in;
DCHECK_EQ(condition->InputAt(0)->GetType(), Primitive::kPrimLong);
- GenerateLongComparesAndJumps(condition, true_target, false_target, codegen_);
+ GenerateLongComparesAndJumps(condition, true_target, false_target, codegen_, is_far_target);
if (false_target != &fallthrough) {
__ B(false_target);
@@ -3057,7 +3059,7 @@
// the HCondition, generate the comparison directly.
Primitive::Type type = condition->InputAt(0)->GetType();
if (type == Primitive::kPrimLong || Primitive::IsFloatingPointType(type)) {
- GenerateCompareTestAndBranch(condition, true_target, false_target);
+ GenerateCompareTestAndBranch(condition, true_target, false_target, far_target);
return;
}
@@ -3076,14 +3078,14 @@
if (right.IsImmediate() && right.GetImmediate() == 0 && (arm_cond.Is(ne) || arm_cond.Is(eq))) {
if (arm_cond.Is(eq)) {
- __ CompareAndBranchIfZero(left, non_fallthrough_target);
+ __ CompareAndBranchIfZero(left, non_fallthrough_target, far_target);
} else {
DCHECK(arm_cond.Is(ne));
- __ CompareAndBranchIfNonZero(left, non_fallthrough_target);
+ __ CompareAndBranchIfNonZero(left, non_fallthrough_target, far_target);
}
} else {
__ Cmp(left, right);
- __ B(arm_cond, non_fallthrough_target);
+ __ B(arm_cond, non_fallthrough_target, far_target);
}
}
diff --git a/compiler/optimizing/code_generator_arm_vixl.h b/compiler/optimizing/code_generator_arm_vixl.h
index 91f7524..91e9a3ed 100644
--- a/compiler/optimizing/code_generator_arm_vixl.h
+++ b/compiler/optimizing/code_generator_arm_vixl.h
@@ -400,7 +400,8 @@
bool far_target = true);
void GenerateCompareTestAndBranch(HCondition* condition,
vixl::aarch32::Label* true_target,
- vixl::aarch32::Label* false_target);
+ vixl::aarch32::Label* false_target,
+ bool is_far_target = true);
void DivRemOneOrMinusOne(HBinaryOperation* instruction);
void DivRemByPowerOfTwo(HBinaryOperation* instruction);
void GenerateDivRemWithAnyConstant(HBinaryOperation* instruction);
diff --git a/compiler/optimizing/intrinsics_arm.cc b/compiler/optimizing/intrinsics_arm.cc
index e8a62aa..9803c9a 100644
--- a/compiler/optimizing/intrinsics_arm.cc
+++ b/compiler/optimizing/intrinsics_arm.cc
@@ -2758,12 +2758,15 @@
int32_t offset = Thread::InterruptedOffset<kArmPointerSize>().Int32Value();
__ LoadFromOffset(kLoadWord, out, TR, offset);
Label done;
- __ CompareAndBranchIfZero(out, &done);
+ Label* const final_label = codegen_->GetFinalLabel(invoke, &done);
+ __ CompareAndBranchIfZero(out, final_label);
__ dmb(ISH);
__ LoadImmediate(IP, 0);
__ StoreToOffset(kStoreWord, IP, TR, offset);
__ dmb(ISH);
- __ Bind(&done);
+ if (done.IsLinked()) {
+ __ Bind(&done);
+ }
}
UNIMPLEMENTED_INTRINSIC(ARM, MathMinDoubleDouble)
diff --git a/compiler/optimizing/intrinsics_arm_vixl.cc b/compiler/optimizing/intrinsics_arm_vixl.cc
index ce3ba52..1a33b0e 100644
--- a/compiler/optimizing/intrinsics_arm_vixl.cc
+++ b/compiler/optimizing/intrinsics_arm_vixl.cc
@@ -3127,7 +3127,7 @@
__ Add(out, in, -info.low);
__ Cmp(out, info.high - info.low + 1);
vixl32::Label allocate, done;
- __ B(hs, &allocate);
+ __ B(hs, &allocate, /* is_far_target */ false);
// If the value is within the bounds, load the j.l.Integer directly from the array.
uint32_t data_offset = mirror::Array::DataOffset(kHeapReferenceSize).Uint32Value();
uint32_t address = dchecked_integral_cast<uint32_t>(reinterpret_cast<uintptr_t>(info.cache));
@@ -3164,12 +3164,15 @@
UseScratchRegisterScope temps(assembler->GetVIXLAssembler());
vixl32::Register temp = temps.Acquire();
vixl32::Label done;
- __ CompareAndBranchIfZero(out, &done, /* far_target */ false);
+ vixl32::Label* const final_label = codegen_->GetFinalLabel(invoke, &done);
+ __ CompareAndBranchIfZero(out, final_label, /* far_target */ false);
__ Dmb(vixl32::ISH);
__ Mov(temp, 0);
assembler->StoreToOffset(kStoreWord, temp, tr, offset);
__ Dmb(vixl32::ISH);
- __ Bind(&done);
+ if (done.IsReferenced()) {
+ __ Bind(&done);
+ }
}
UNIMPLEMENTED_INTRINSIC(ARMVIXL, MathRoundDouble) // Could be done by changing rounding mode, maybe?