ARM/ARM64: Use trampolines for slow-path entrypoint calls.
This reduces the size of the generated code. We do this only
for AOT compilation where we get the most benefit.
Sizes of aosp_taimen-userdebug prebuilts:
- before:
arm/boot*.oat: 19624804
arm64/boot*.oat: 23265752
oat/arm64/services.odex: 22417968
- after:
arm/boot*.oat: 19460500 (-160KiB)
arm64/boot*.oat: 22957928 (-301KiB)
oat/arm64/services.odex: 21957864 (-449KiB)
Test: m test-art-host-gtest
Test: aosp_taimen-userdebug boots.
Test: run-gtests.sh
Test: testrunner.py --target --optimizing
Bug: 12607709
Change-Id: Ie9dbd1ba256173e4e439e8bbb8832a791965cbe6
diff --git a/compiler/optimizing/code_generator_arm_vixl.cc b/compiler/optimizing/code_generator_arm_vixl.cc
index 19d04c9..ac09183 100644
--- a/compiler/optimizing/code_generator_arm_vixl.cc
+++ b/compiler/optimizing/code_generator_arm_vixl.cc
@@ -1856,8 +1856,6 @@
instruction_visitor_(graph, this),
move_resolver_(graph->GetAllocator(), this),
assembler_(graph->GetAllocator()),
- uint32_literals_(std::less<uint32_t>(),
- graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
boot_image_method_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
method_bss_entry_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
boot_image_type_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
@@ -1865,7 +1863,10 @@
boot_image_string_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
string_bss_entry_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
boot_image_intrinsic_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
+ call_entrypoint_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
baker_read_barrier_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
+ uint32_literals_(std::less<uint32_t>(),
+ graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
jit_string_patches_(StringReferenceValueComparator(),
graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
jit_class_patches_(TypeReferenceValueComparator(),
@@ -2383,15 +2384,31 @@
uint32_t dex_pc,
SlowPathCode* slow_path) {
ValidateInvokeRuntime(entrypoint, instruction, slow_path);
- __ Ldr(lr, MemOperand(tr, GetThreadOffset<kArmPointerSize>(entrypoint).Int32Value()));
- // Ensure the pc position is recorded immediately after the `blx` instruction.
- // blx in T32 has only 16bit encoding that's why a stricter check for the scope is used.
- ExactAssemblyScope aas(GetVIXLAssembler(),
- vixl32::k16BitT32InstructionSizeInBytes,
- CodeBufferCheckScope::kExactSize);
- __ blx(lr);
- if (EntrypointRequiresStackMap(entrypoint)) {
- RecordPcInfo(instruction, dex_pc, slow_path);
+
+ ThreadOffset32 entrypoint_offset = GetThreadOffset<kArmPointerSize>(entrypoint);
+ // Reduce code size for AOT by using shared trampolines for slow path runtime calls across the
+ // entire oat file. This adds an extra branch and we do not want to slow down the main path.
+ // For JIT, thunk sharing is per-method, so the gains would be smaller or even negative.
+ if (slow_path == nullptr || Runtime::Current()->UseJitCompilation()) {
+ __ Ldr(lr, MemOperand(tr, entrypoint_offset.Int32Value()));
+ // Ensure the pc position is recorded immediately after the `blx` instruction.
+ // blx in T32 has only 16bit encoding that's why a stricter check for the scope is used.
+ ExactAssemblyScope aas(GetVIXLAssembler(),
+ vixl32::k16BitT32InstructionSizeInBytes,
+ CodeBufferCheckScope::kExactSize);
+ __ blx(lr);
+ if (EntrypointRequiresStackMap(entrypoint)) {
+ RecordPcInfo(instruction, dex_pc, slow_path);
+ }
+ } else {
+ // Ensure the pc position is recorded immediately after the `bl` instruction.
+ ExactAssemblyScope aas(GetVIXLAssembler(),
+ vixl32::k32BitT32InstructionSizeInBytes,
+ CodeBufferCheckScope::kExactSize);
+ EmitEntrypointThunkCall(entrypoint_offset);
+ if (EntrypointRequiresStackMap(entrypoint)) {
+ RecordPcInfo(instruction, dex_pc, slow_path);
+ }
}
}
@@ -8858,6 +8875,17 @@
return &patches->back();
}
+void CodeGeneratorARMVIXL::EmitEntrypointThunkCall(ThreadOffset32 entrypoint_offset) {
+ DCHECK(!__ AllowMacroInstructions()); // In ExactAssemblyScope.
+ DCHECK(!Runtime::Current()->UseJitCompilation());
+ call_entrypoint_patches_.emplace_back(/*dex_file*/ nullptr, entrypoint_offset.Uint32Value());
+ vixl::aarch32::Label* bl_label = &call_entrypoint_patches_.back().label;
+ __ bind(bl_label);
+ vixl32::Label placeholder_label;
+ __ bl(&placeholder_label); // Placeholder, patched at link-time.
+ __ bind(&placeholder_label);
+}
+
void CodeGeneratorARMVIXL::EmitBakerReadBarrierBne(uint32_t custom_data) {
DCHECK(!__ AllowMacroInstructions()); // In ExactAssemblyScope.
if (Runtime::Current()->UseJitCompilation()) {
@@ -8980,6 +9008,7 @@
/* MOVW+MOVT for each entry */ 2u * boot_image_string_patches_.size() +
/* MOVW+MOVT for each entry */ 2u * string_bss_entry_patches_.size() +
/* MOVW+MOVT for each entry */ 2u * boot_image_intrinsic_patches_.size() +
+ call_entrypoint_patches_.size() +
baker_read_barrier_patches_.size();
linker_patches->reserve(size);
if (GetCompilerOptions().IsBootImage()) {
@@ -9004,6 +9033,11 @@
type_bss_entry_patches_, linker_patches);
EmitPcRelativeLinkerPatches<linker::LinkerPatch::StringBssEntryPatch>(
string_bss_entry_patches_, linker_patches);
+ for (const PatchInfo<vixl32::Label>& info : call_entrypoint_patches_) {
+ DCHECK(info.target_dex_file == nullptr);
+ linker_patches->push_back(linker::LinkerPatch::CallEntrypointPatch(
+ info.label.GetLocation(), info.offset_or_index));
+ }
for (const BakerReadBarrierPatchInfo& info : baker_read_barrier_patches_) {
linker_patches->push_back(linker::LinkerPatch::BakerReadBarrierBranchPatch(
info.label.GetLocation(), info.custom_data));
@@ -9012,7 +9046,8 @@
}
bool CodeGeneratorARMVIXL::NeedsThunkCode(const linker::LinkerPatch& patch) const {
- return patch.GetType() == linker::LinkerPatch::Type::kBakerReadBarrierBranch ||
+ return patch.GetType() == linker::LinkerPatch::Type::kCallEntrypoint ||
+ patch.GetType() == linker::LinkerPatch::Type::kBakerReadBarrierBranch ||
patch.GetType() == linker::LinkerPatch::Type::kCallRelative;
}
@@ -9021,23 +9056,30 @@
/*out*/ std::string* debug_name) {
arm::ArmVIXLAssembler assembler(GetGraph()->GetAllocator());
switch (patch.GetType()) {
- case linker::LinkerPatch::Type::kCallRelative:
+ case linker::LinkerPatch::Type::kCallRelative: {
// The thunk just uses the entry point in the ArtMethod. This works even for calls
// to the generic JNI and interpreter trampolines.
- assembler.LoadFromOffset(
- arm::kLoadWord,
- vixl32::pc,
- vixl32::r0,
- ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize).Int32Value());
+ MemberOffset offset = ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize);
+ assembler.LoadFromOffset(arm::kLoadWord, vixl32::pc, vixl32::r0, offset.Int32Value());
assembler.GetVIXLAssembler()->Bkpt(0);
if (GetCompilerOptions().GenerateAnyDebugInfo()) {
*debug_name = "MethodCallThunk";
}
break;
- case linker::LinkerPatch::Type::kBakerReadBarrierBranch:
+ }
+ case linker::LinkerPatch::Type::kCallEntrypoint: {
+ assembler.LoadFromOffset(arm::kLoadWord, vixl32::pc, tr, patch.EntrypointOffset());
+ assembler.GetVIXLAssembler()->Bkpt(0);
+ if (GetCompilerOptions().GenerateAnyDebugInfo()) {
+ *debug_name = "EntrypointCallThunk_" + std::to_string(patch.EntrypointOffset());
+ }
+ break;
+ }
+ case linker::LinkerPatch::Type::kBakerReadBarrierBranch: {
DCHECK_EQ(patch.GetBakerCustomValue2(), 0u);
CompileBakerReadBarrierThunk(assembler, patch.GetBakerCustomValue1(), debug_name);
break;
+ }
default:
LOG(FATAL) << "Unexpected patch type " << patch.GetType();
UNREACHABLE();