Compiler changes for bitstring based type checks.
We guard the use of this feature with a compile-time flag,
set to true in this CL.
Boot image size for aosp_taimen-userdebug in AOSP master:
- before:
arm boot*.oat: 63604740
arm64 boot*.oat: 74237864
- after:
arm boot*.oat: 63531172 (-72KiB, -0.1%)
arm64 boot*.oat: 74135008 (-100KiB, -0.1%)
The new TypeCheckBenchmark yields the following changes
using the little cores of taimen fixed at 1.4016GHz:
32-bit 64-bit
timeCheckCastLevel1ToLevel1 11.48->15.80 11.47->15.78
timeCheckCastLevel2ToLevel1 15.08->15.79 15.08->15.79
timeCheckCastLevel3ToLevel1 19.01->15.82 17.94->15.81
timeCheckCastLevel9ToLevel1 42.55->15.79 42.63->15.81
timeCheckCastLevel9ToLevel2 39.70->14.36 39.70->14.35
timeInstanceOfLevel1ToLevel1 13.74->17.93 13.76->17.95
timeInstanceOfLevel2ToLevel1 17.02->17.95 16.99->17.93
timeInstanceOfLevel3ToLevel1 24.03->17.95 24.45->17.95
timeInstanceOfLevel9ToLevel1 47.13->17.95 47.14->18.00
timeInstanceOfLevel9ToLevel2 44.19->16.52 44.27->16.51
This suggests that the bitstring typecheck should not be
used for exact type checks which would be equivalent to the
"Level1ToLevel1" benchmark. Whether the implementation is
a beneficial replacement for the kClassHierarchyCheck and
kAbstractClassCheck on average depends on how many levels
from the target class (or Object for a negative result) is
a typical object's class.
Test: m test-art-host-gtest
Test: testrunner.py --host --optimizing --jit
Test: testrunner.py --host -t 670-bitstring-type-check
Test: Pixel 2 XL boots.
Test: testrunner.py --target --optimizing --jit
Test: testrunner.py --target -t 670-bitstring-type-check
Bug: 64692057
Bug: 71853552
Bug: 26687569
Change-Id: I538d7e036b5a8ae2cc3fe77662a5903d74854562
diff --git a/compiler/optimizing/code_generator_mips64.cc b/compiler/optimizing/code_generator_mips64.cc
index bcfe051..3ae8a30 100644
--- a/compiler/optimizing/code_generator_mips64.cc
+++ b/compiler/optimizing/code_generator_mips64.cc
@@ -1775,6 +1775,34 @@
__ Bind(slow_path->GetExitLabel());
}
+void InstructionCodeGeneratorMIPS64::GenerateBitstringTypeCheckCompare(HTypeCheckInstruction* check,
+ GpuRegister temp) {
+ uint32_t path_to_root = check->GetBitstringPathToRoot();
+ uint32_t mask = check->GetBitstringMask();
+ DCHECK(IsPowerOfTwo(mask + 1));
+ size_t mask_bits = WhichPowerOf2(mask + 1);
+
+ if (mask_bits == 16u) {
+ // Load only the bitstring part of the status word.
+ __ LoadFromOffset(
+ kLoadUnsignedHalfword, temp, temp, mirror::Class::StatusOffset().Int32Value());
+ // Compare the bitstring bits using XOR.
+ __ Xori(temp, temp, dchecked_integral_cast<uint16_t>(path_to_root));
+ } else {
+ // /* uint32_t */ temp = temp->status_
+ __ LoadFromOffset(kLoadWord, temp, temp, mirror::Class::StatusOffset().Int32Value());
+ // Compare the bitstring bits using XOR.
+ if (IsUint<16>(path_to_root)) {
+ __ Xori(temp, temp, dchecked_integral_cast<uint16_t>(path_to_root));
+ } else {
+ __ LoadConst32(TMP, path_to_root);
+ __ Xor(temp, temp, TMP);
+ }
+ // Shift out bits that do not contribute to the comparison.
+ __ Sll(temp, temp, 32 - mask_bits);
+ }
+}
+
void InstructionCodeGeneratorMIPS64::GenerateMemoryBarrier(MemBarrierKind kind ATTRIBUTE_UNUSED) {
__ Sync(0); // only stype 0 is supported
}
@@ -2844,12 +2872,20 @@
case TypeCheckKind::kInterfaceCheck:
call_kind = LocationSummary::kCallOnSlowPath;
break;
+ case TypeCheckKind::kBitstringCheck:
+ break;
}
LocationSummary* locations =
new (GetGraph()->GetAllocator()) LocationSummary(instruction, call_kind);
locations->SetInAt(0, Location::RequiresRegister());
- locations->SetInAt(1, Location::RequiresRegister());
+ if (type_check_kind == TypeCheckKind::kBitstringCheck) {
+ locations->SetInAt(1, Location::ConstantLocation(instruction->InputAt(1)->AsConstant()));
+ locations->SetInAt(2, Location::ConstantLocation(instruction->InputAt(2)->AsConstant()));
+ locations->SetInAt(3, Location::ConstantLocation(instruction->InputAt(3)->AsConstant()));
+ } else {
+ locations->SetInAt(1, Location::RequiresRegister());
+ }
locations->AddRegisterTemps(NumberOfCheckCastTemps(type_check_kind));
}
@@ -2858,7 +2894,7 @@
LocationSummary* locations = instruction->GetLocations();
Location obj_loc = locations->InAt(0);
GpuRegister obj = obj_loc.AsRegister<GpuRegister>();
- GpuRegister cls = locations->InAt(1).AsRegister<GpuRegister>();
+ Location cls = locations->InAt(1);
Location temp_loc = locations->GetTemp(0);
GpuRegister temp = temp_loc.AsRegister<GpuRegister>();
const size_t num_temps = NumberOfCheckCastTemps(type_check_kind);
@@ -2908,7 +2944,7 @@
kWithoutReadBarrier);
// Jump to slow path for throwing the exception or doing a
// more involved array check.
- __ Bnec(temp, cls, slow_path->GetEntryLabel());
+ __ Bnec(temp, cls.AsRegister<GpuRegister>(), slow_path->GetEntryLabel());
break;
}
@@ -2934,7 +2970,7 @@
// exception.
__ Beqzc(temp, slow_path->GetEntryLabel());
// Otherwise, compare the classes.
- __ Bnec(temp, cls, &loop);
+ __ Bnec(temp, cls.AsRegister<GpuRegister>(), &loop);
break;
}
@@ -2949,7 +2985,7 @@
// Walk over the class hierarchy to find a match.
Mips64Label loop;
__ Bind(&loop);
- __ Beqc(temp, cls, &done);
+ __ Beqc(temp, cls.AsRegister<GpuRegister>(), &done);
// /* HeapReference<Class> */ temp = temp->super_class_
GenerateReferenceLoadOneRegister(instruction,
temp_loc,
@@ -2972,7 +3008,7 @@
maybe_temp2_loc,
kWithoutReadBarrier);
// Do an exact check.
- __ Beqc(temp, cls, &done);
+ __ Beqc(temp, cls.AsRegister<GpuRegister>(), &done);
// Otherwise, we need to check that the object's class is a non-primitive array.
// /* HeapReference<Class> */ temp = temp->component_type_
GenerateReferenceLoadOneRegister(instruction,
@@ -3031,7 +3067,21 @@
__ Daddiu(temp, temp, 2 * kHeapReferenceSize);
__ Addiu(TMP, TMP, -2);
// Compare the classes and continue the loop if they do not match.
- __ Bnec(AT, cls, &loop);
+ __ Bnec(AT, cls.AsRegister<GpuRegister>(), &loop);
+ break;
+ }
+
+ case TypeCheckKind::kBitstringCheck: {
+ // /* HeapReference<Class> */ temp = obj->klass_
+ GenerateReferenceLoadTwoRegisters(instruction,
+ temp_loc,
+ obj_loc,
+ class_offset,
+ maybe_temp2_loc,
+ kWithoutReadBarrier);
+
+ GenerateBitstringTypeCheckCompare(instruction, temp);
+ __ Bnezc(temp, slow_path->GetEntryLabel());
break;
}
}
@@ -5524,6 +5574,8 @@
case TypeCheckKind::kInterfaceCheck:
call_kind = LocationSummary::kCallOnSlowPath;
break;
+ case TypeCheckKind::kBitstringCheck:
+ break;
}
LocationSummary* locations =
@@ -5532,7 +5584,13 @@
locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers.
}
locations->SetInAt(0, Location::RequiresRegister());
- locations->SetInAt(1, Location::RequiresRegister());
+ if (type_check_kind == TypeCheckKind::kBitstringCheck) {
+ locations->SetInAt(1, Location::ConstantLocation(instruction->InputAt(1)->AsConstant()));
+ locations->SetInAt(2, Location::ConstantLocation(instruction->InputAt(2)->AsConstant()));
+ locations->SetInAt(3, Location::ConstantLocation(instruction->InputAt(3)->AsConstant()));
+ } else {
+ locations->SetInAt(1, Location::RequiresRegister());
+ }
// The output does overlap inputs.
// Note that TypeCheckSlowPathMIPS64 uses this register too.
locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
@@ -5544,7 +5602,7 @@
LocationSummary* locations = instruction->GetLocations();
Location obj_loc = locations->InAt(0);
GpuRegister obj = obj_loc.AsRegister<GpuRegister>();
- GpuRegister cls = locations->InAt(1).AsRegister<GpuRegister>();
+ Location cls = locations->InAt(1);
Location out_loc = locations->Out();
GpuRegister out = out_loc.AsRegister<GpuRegister>();
const size_t num_temps = NumberOfInstanceOfTemps(type_check_kind);
@@ -5574,7 +5632,7 @@
maybe_temp_loc,
kCompilerReadBarrierOption);
// Classes must be equal for the instanceof to succeed.
- __ Xor(out, out, cls);
+ __ Xor(out, out, cls.AsRegister<GpuRegister>());
__ Sltiu(out, out, 1);
break;
}
@@ -5599,7 +5657,7 @@
kCompilerReadBarrierOption);
// If `out` is null, we use it for the result, and jump to `done`.
__ Beqzc(out, &done);
- __ Bnec(out, cls, &loop);
+ __ Bnec(out, cls.AsRegister<GpuRegister>(), &loop);
__ LoadConst32(out, 1);
break;
}
@@ -5615,7 +5673,7 @@
// Walk over the class hierarchy to find a match.
Mips64Label loop, success;
__ Bind(&loop);
- __ Beqc(out, cls, &success);
+ __ Beqc(out, cls.AsRegister<GpuRegister>(), &success);
// /* HeapReference<Class> */ out = out->super_class_
GenerateReferenceLoadOneRegister(instruction,
out_loc,
@@ -5640,7 +5698,7 @@
kCompilerReadBarrierOption);
// Do an exact check.
Mips64Label success;
- __ Beqc(out, cls, &success);
+ __ Beqc(out, cls.AsRegister<GpuRegister>(), &success);
// Otherwise, we need to check that the object's class is a non-primitive array.
// /* HeapReference<Class> */ out = out->component_type_
GenerateReferenceLoadOneRegister(instruction,
@@ -5672,7 +5730,7 @@
slow_path = new (codegen_->GetScopedAllocator()) TypeCheckSlowPathMIPS64(
instruction, /* is_fatal */ false);
codegen_->AddSlowPath(slow_path);
- __ Bnec(out, cls, slow_path->GetEntryLabel());
+ __ Bnec(out, cls.AsRegister<GpuRegister>(), slow_path->GetEntryLabel());
__ LoadConst32(out, 1);
break;
}
@@ -5704,6 +5762,20 @@
__ Bc(slow_path->GetEntryLabel());
break;
}
+
+ case TypeCheckKind::kBitstringCheck: {
+ // /* HeapReference<Class> */ temp = obj->klass_
+ GenerateReferenceLoadTwoRegisters(instruction,
+ out_loc,
+ obj_loc,
+ class_offset,
+ maybe_temp_loc,
+ kWithoutReadBarrier);
+
+ GenerateBitstringTypeCheckCompare(instruction, out);
+ __ Sltiu(out, out, 1);
+ break;
+ }
}
__ Bind(&done);