Align method code rather than method header in oat files.
This has always been the intent. For example, the function
name CompiledMethod::AlignCode() shows this intent.
Test: Run ART test suite on host and Nexus 9.
Change-Id: I11ae8963fc537fee853fe82e3aca22e77907eae7
diff --git a/compiler/linker/arm/relative_patcher_arm_base.cc b/compiler/linker/arm/relative_patcher_arm_base.cc
index d4dd978..2471f79 100644
--- a/compiler/linker/arm/relative_patcher_arm_base.cc
+++ b/compiler/linker/arm/relative_patcher_arm_base.cc
@@ -31,10 +31,6 @@
}
uint32_t ArmBaseRelativePatcher::ReserveSpaceEnd(uint32_t offset) {
- // NOTE: The final thunk can be reserved from InitCodeMethodVisitor::EndClass() while it
- // may be written early by WriteCodeMethodVisitor::VisitMethod() for a deduplicated chunk
- // of code. To avoid any alignment discrepancies for the final chunk, we always align the
- // offset after reserving of writing any chunk.
uint32_t aligned_offset = CompiledMethod::AlignCode(offset, instruction_set_);
bool needs_thunk = ReserveSpaceProcessPatches(aligned_offset,
MethodReference(nullptr, 0u),
@@ -46,7 +42,7 @@
unprocessed_patches_.clear();
thunk_locations_.push_back(aligned_offset);
- offset = CompiledMethod::AlignCode(aligned_offset + thunk_code_.size(), instruction_set_);
+ offset = aligned_offset + thunk_code_.size();
}
return offset;
}
@@ -65,13 +61,7 @@
if (UNLIKELY(!WriteRelCallThunk(out, ArrayRef<const uint8_t>(thunk_code_)))) {
return 0u;
}
- uint32_t thunk_end_offset = aligned_offset + thunk_code_.size();
- // Align after writing chunk, see the ReserveSpace() above.
- offset = CompiledMethod::AlignCode(thunk_end_offset, instruction_set_);
- aligned_code_delta = offset - thunk_end_offset;
- if (aligned_code_delta != 0u && !WriteCodeAlignment(out, aligned_code_delta)) {
- return 0u;
- }
+ offset = aligned_offset + thunk_code_.size();
}
return offset;
}
@@ -92,7 +82,7 @@
MethodReference method_ref,
uint32_t max_extra_space) {
uint32_t quick_code_size = compiled_method->GetQuickCode().size();
- uint32_t quick_code_offset = compiled_method->AlignCode(offset) + sizeof(OatQuickMethodHeader);
+ uint32_t quick_code_offset = compiled_method->AlignCode(offset + sizeof(OatQuickMethodHeader));
uint32_t next_aligned_offset = compiled_method->AlignCode(quick_code_offset + quick_code_size);
// Adjust for extra space required by the subclass.
next_aligned_offset = compiled_method->AlignCode(next_aligned_offset + max_extra_space);
@@ -106,9 +96,9 @@
if (needs_thunk) {
// A single thunk will cover all pending patches.
unprocessed_patches_.clear();
- uint32_t thunk_location = compiled_method->AlignCode(offset);
+ uint32_t thunk_location = CompiledMethod::AlignCode(offset, instruction_set_);
thunk_locations_.push_back(thunk_location);
- offset = CompiledMethod::AlignCode(thunk_location + thunk_code_.size(), instruction_set_);
+ offset = thunk_location + thunk_code_.size();
}
}
for (const LinkerPatch& patch : compiled_method->GetPatches()) {
diff --git a/compiler/linker/arm/relative_patcher_thumb2_test.cc b/compiler/linker/arm/relative_patcher_thumb2_test.cc
index a8078e3..eace3d4 100644
--- a/compiler/linker/arm/relative_patcher_thumb2_test.cc
+++ b/compiler/linker/arm/relative_patcher_thumb2_test.cc
@@ -48,18 +48,18 @@
const ArrayRef<const LinkerPatch>& method3_patches,
uint32_t distance_without_thunks) {
CHECK_EQ(distance_without_thunks % kArmAlignment, 0u);
- const uint32_t method1_offset =
- CompiledCode::AlignCode(kTrampolineSize, kThumb2) + sizeof(OatQuickMethodHeader);
+ uint32_t method1_offset =
+ kTrampolineSize + CodeAlignmentSize(kTrampolineSize) + sizeof(OatQuickMethodHeader);
AddCompiledMethod(MethodRef(1u), method1_code, method1_patches);
// We want to put the method3 at a very precise offset.
const uint32_t method3_offset = method1_offset + distance_without_thunks;
- CHECK_ALIGNED(method3_offset - sizeof(OatQuickMethodHeader), kArmAlignment);
+ CHECK_ALIGNED(method3_offset, kArmAlignment);
// Calculate size of method2 so that we put method3 at the correct place.
+ const uint32_t method1_end = method1_offset + method1_code.size();
const uint32_t method2_offset =
- CompiledCode::AlignCode(method1_offset + method1_code.size(), kThumb2) +
- sizeof(OatQuickMethodHeader);
+ method1_end + CodeAlignmentSize(method1_end) + sizeof(OatQuickMethodHeader);
const uint32_t method2_size = (method3_offset - sizeof(OatQuickMethodHeader) - method2_offset);
std::vector<uint8_t> method2_raw_code(method2_size);
ArrayRef<const uint8_t> method2_code(method2_raw_code);
@@ -78,8 +78,11 @@
if (result3.second == method3_offset + 1 /* thumb mode */) {
return false; // No thunk.
} else {
- uint32_t aligned_thunk_size = CompiledCode::AlignCode(ThunkSize(), kThumb2);
- CHECK_EQ(result3.second, method3_offset + aligned_thunk_size + 1 /* thumb mode */);
+ uint32_t thunk_end =
+ CompiledCode::AlignCode(method3_offset - sizeof(OatQuickMethodHeader), kThumb2) +
+ ThunkSize();
+ uint32_t header_offset = thunk_end + CodeAlignmentSize(thunk_end);
+ CHECK_EQ(result3.second, header_offset + sizeof(OatQuickMethodHeader) + 1 /* thumb mode */);
return true; // Thunk present.
}
}
@@ -352,9 +355,12 @@
uint32_t method1_offset = GetMethodOffset(1u);
uint32_t method3_offset = GetMethodOffset(3u);
+ ASSERT_TRUE(IsAligned<kArmAlignment>(method3_offset));
uint32_t method3_header_offset = method3_offset - sizeof(OatQuickMethodHeader);
- ASSERT_TRUE(IsAligned<kArmAlignment>(method3_header_offset));
- uint32_t thunk_offset = method3_header_offset - CompiledCode::AlignCode(ThunkSize(), kThumb2);
+ uint32_t thunk_offset =
+ RoundDown(method3_header_offset - ThunkSize(), GetInstructionSetAlignment(kThumb2));
+ DCHECK_EQ(thunk_offset + ThunkSize() + CodeAlignmentSize(thunk_offset + ThunkSize()),
+ method3_header_offset);
ASSERT_TRUE(IsAligned<kArmAlignment>(thunk_offset));
uint32_t diff = thunk_offset - (method1_offset + bl_offset_in_method1 + 4u /* PC adjustment */);
ASSERT_EQ(diff & 1u, 0u);
diff --git a/compiler/linker/arm64/relative_patcher_arm64.cc b/compiler/linker/arm64/relative_patcher_arm64.cc
index fdd14be..4c8788e 100644
--- a/compiler/linker/arm64/relative_patcher_arm64.cc
+++ b/compiler/linker/arm64/relative_patcher_arm64.cc
@@ -83,7 +83,7 @@
// Now that we have the actual offset where the code will be placed, locate the ADRP insns
// that actually require the thunk.
- uint32_t quick_code_offset = compiled_method->AlignCode(offset) + sizeof(OatQuickMethodHeader);
+ uint32_t quick_code_offset = compiled_method->AlignCode(offset + sizeof(OatQuickMethodHeader));
ArrayRef<const uint8_t> code = compiled_method->GetQuickCode();
uint32_t thunk_offset = compiled_method->AlignCode(quick_code_offset + code.size());
DCHECK(compiled_method != nullptr);
diff --git a/compiler/linker/arm64/relative_patcher_arm64_test.cc b/compiler/linker/arm64/relative_patcher_arm64_test.cc
index 09729fd..573de73 100644
--- a/compiler/linker/arm64/relative_patcher_arm64_test.cc
+++ b/compiler/linker/arm64/relative_patcher_arm64_test.cc
@@ -67,36 +67,39 @@
const ArrayRef<const LinkerPatch>& last_method_patches,
uint32_t distance_without_thunks) {
CHECK_EQ(distance_without_thunks % kArm64Alignment, 0u);
- const uint32_t method1_offset =
- CompiledCode::AlignCode(kTrampolineSize, kArm64) + sizeof(OatQuickMethodHeader);
+ uint32_t method1_offset =
+ kTrampolineSize + CodeAlignmentSize(kTrampolineSize) + sizeof(OatQuickMethodHeader);
AddCompiledMethod(MethodRef(1u), method1_code, method1_patches);
- const uint32_t gap_start =
- CompiledCode::AlignCode(method1_offset + method1_code.size(), kArm64);
+ const uint32_t gap_start = method1_offset + method1_code.size();
// We want to put the method3 at a very precise offset.
const uint32_t last_method_offset = method1_offset + distance_without_thunks;
+ CHECK_ALIGNED(last_method_offset, kArm64Alignment);
const uint32_t gap_end = last_method_offset - sizeof(OatQuickMethodHeader);
- CHECK_ALIGNED(gap_end, kArm64Alignment);
- // Fill the gap with intermediate methods in chunks of 2MiB and the last in [2MiB, 4MiB).
+ // Fill the gap with intermediate methods in chunks of 2MiB and the first in [2MiB, 4MiB).
// (This allows deduplicating the small chunks to avoid using 256MiB of memory for +-128MiB
- // offsets by this test.)
+ // offsets by this test. Making the first chunk bigger makes it easy to give all intermediate
+ // methods the same alignment of the end, so the thunk insertion adds a predictable size as
+ // long as it's after the first chunk.)
uint32_t method_idx = 2u;
constexpr uint32_t kSmallChunkSize = 2 * MB;
std::vector<uint8_t> gap_code;
- size_t gap_size = gap_end - gap_start;
- for (; gap_size >= 2u * kSmallChunkSize; gap_size -= kSmallChunkSize) {
- uint32_t chunk_code_size = kSmallChunkSize - sizeof(OatQuickMethodHeader);
+ uint32_t gap_size = gap_end - gap_start;
+ uint32_t num_small_chunks = std::max(gap_size / kSmallChunkSize, 1u) - 1u;
+ uint32_t chunk_start = gap_start;
+ uint32_t chunk_size = gap_size - num_small_chunks * kSmallChunkSize;
+ for (uint32_t i = 0; i <= num_small_chunks; ++i) { // num_small_chunks+1 iterations.
+ uint32_t chunk_code_size =
+ chunk_size - CodeAlignmentSize(chunk_start) - sizeof(OatQuickMethodHeader);
gap_code.resize(chunk_code_size, 0u);
AddCompiledMethod(MethodRef(method_idx), ArrayRef<const uint8_t>(gap_code),
ArrayRef<const LinkerPatch>());
method_idx += 1u;
+ chunk_start += chunk_size;
+ chunk_size = kSmallChunkSize; // For all but the first chunk.
+ DCHECK_EQ(CodeAlignmentSize(gap_end), CodeAlignmentSize(chunk_start));
}
- uint32_t chunk_code_size = gap_size - sizeof(OatQuickMethodHeader);
- gap_code.resize(chunk_code_size, 0u);
- AddCompiledMethod(MethodRef(method_idx), ArrayRef<const uint8_t>(gap_code),
- ArrayRef<const LinkerPatch>());
- method_idx += 1u;
// Add the last method and link
AddCompiledMethod(MethodRef(method_idx), last_method_code, last_method_patches);
@@ -109,8 +112,9 @@
// There may be a thunk before method2.
if (last_result.second != last_method_offset) {
// Thunk present. Check that there's only one.
- uint32_t aligned_thunk_size = CompiledCode::AlignCode(ThunkSize(), kArm64);
- CHECK_EQ(last_result.second, last_method_offset + aligned_thunk_size);
+ uint32_t thunk_end = CompiledCode::AlignCode(gap_end, kArm64) + ThunkSize();
+ uint32_t header_offset = thunk_end + CodeAlignmentSize(thunk_end);
+ CHECK_EQ(last_result.second, header_offset + sizeof(OatQuickMethodHeader));
}
return method_idx;
}
@@ -341,7 +345,7 @@
uint32_t dex_cache_arrays_begin,
uint32_t element_offset) {
uint32_t method1_offset =
- CompiledCode::AlignCode(kTrampolineSize, kArm64) + sizeof(OatQuickMethodHeader);
+ kTrampolineSize + CodeAlignmentSize(kTrampolineSize) + sizeof(OatQuickMethodHeader);
ASSERT_LT(method1_offset, adrp_offset);
CHECK_ALIGNED(adrp_offset, 4u);
uint32_t num_nops = (adrp_offset - method1_offset) / 4u;
@@ -391,7 +395,7 @@
bool has_thunk,
uint32_t string_offset) {
uint32_t method1_offset =
- CompiledCode::AlignCode(kTrampolineSize, kArm64) + sizeof(OatQuickMethodHeader);
+ kTrampolineSize + CodeAlignmentSize(kTrampolineSize) + sizeof(OatQuickMethodHeader);
ASSERT_LT(method1_offset, adrp_offset);
CHECK_ALIGNED(adrp_offset, 4u);
uint32_t num_nops = (adrp_offset - method1_offset) / 4u;
@@ -614,10 +618,12 @@
uint32_t method1_offset = GetMethodOffset(1u);
uint32_t last_method_offset = GetMethodOffset(last_method_idx);
+ ASSERT_TRUE(IsAligned<kArm64Alignment>(last_method_offset));
uint32_t last_method_header_offset = last_method_offset - sizeof(OatQuickMethodHeader);
- ASSERT_TRUE(IsAligned<kArm64Alignment>(last_method_header_offset));
- uint32_t thunk_offset = last_method_header_offset - CompiledCode::AlignCode(ThunkSize(), kArm64);
- ASSERT_TRUE(IsAligned<kArm64Alignment>(thunk_offset));
+ uint32_t thunk_offset =
+ RoundDown(last_method_header_offset - ThunkSize(), GetInstructionSetAlignment(kArm64));
+ DCHECK_EQ(thunk_offset + ThunkSize() + CodeAlignmentSize(thunk_offset + ThunkSize()),
+ last_method_header_offset);
uint32_t diff = thunk_offset - (method1_offset + bl_offset_in_method1);
CHECK_ALIGNED(diff, 4u);
ASSERT_LT(diff, 128 * MB);
diff --git a/compiler/linker/relative_patcher_test.h b/compiler/linker/relative_patcher_test.h
index ec69107..d21f33e 100644
--- a/compiler/linker/relative_patcher_test.h
+++ b/compiler/linker/relative_patcher_test.h
@@ -98,6 +98,14 @@
patches));
}
+ uint32_t CodeAlignmentSize(uint32_t header_offset_to_align) {
+ // We want to align the code rather than the preheader.
+ uint32_t unaligned_code_offset = header_offset_to_align + sizeof(OatQuickMethodHeader);
+ uint32_t aligned_code_offset =
+ CompiledMethod::AlignCode(unaligned_code_offset, instruction_set_);
+ return aligned_code_offset - unaligned_code_offset;
+ }
+
void Link() {
// Reserve space.
static_assert(kTrampolineOffset == 0u, "Unexpected trampoline offset.");
@@ -106,9 +114,8 @@
for (auto& compiled_method : compiled_methods_) {
offset = patcher_->ReserveSpace(offset, compiled_method.get(), compiled_method_refs_[idx]);
- uint32_t aligned_offset = compiled_method->AlignCode(offset);
- uint32_t aligned_code_delta = aligned_offset - offset;
- offset += aligned_code_delta;
+ uint32_t alignment_size = CodeAlignmentSize(offset);
+ offset += alignment_size;
offset += sizeof(OatQuickMethodHeader);
uint32_t quick_code_offset = offset + compiled_method->CodeDelta();
@@ -136,11 +143,10 @@
for (auto& compiled_method : compiled_methods_) {
offset = patcher_->WriteThunks(&out_, offset);
- uint32_t aligned_offset = compiled_method->AlignCode(offset);
- uint32_t aligned_code_delta = aligned_offset - offset;
- CHECK_LE(aligned_code_delta, sizeof(kPadding));
- out_.WriteFully(kPadding, aligned_code_delta);
- offset += aligned_code_delta;
+ uint32_t alignment_size = CodeAlignmentSize(offset);
+ CHECK_LE(alignment_size, sizeof(kPadding));
+ out_.WriteFully(kPadding, alignment_size);
+ offset += alignment_size;
out_.WriteFully(dummy_header, sizeof(OatQuickMethodHeader));
offset += sizeof(OatQuickMethodHeader);
diff --git a/compiler/oat_writer.cc b/compiler/oat_writer.cc
index b32199f..96b199a 100644
--- a/compiler/oat_writer.cc
+++ b/compiler/oat_writer.cc
@@ -86,6 +86,13 @@
OatHeader* const oat_header_;
};
+inline uint32_t CodeAlignmentSize(uint32_t header_offset, const CompiledMethod& compiled_method) {
+ // We want to align the code rather than the preheader.
+ uint32_t unaligned_code_offset = header_offset + sizeof(OatQuickMethodHeader);
+ uint32_t aligned_code_offset = compiled_method.AlignCode(unaligned_code_offset);
+ return aligned_code_offset - unaligned_code_offset;
+}
+
} // anonymous namespace
// Defines the location of the raw dex file to write.
@@ -816,8 +823,8 @@
uint32_t thumb_offset) {
offset_ = writer_->relative_patcher_->ReserveSpace(
offset_, compiled_method, MethodReference(dex_file_, it.GetMemberIndex()));
- offset_ = compiled_method->AlignCode(offset_);
- DCHECK_ALIGNED_PARAM(offset_,
+ offset_ += CodeAlignmentSize(offset_, *compiled_method);
+ DCHECK_ALIGNED_PARAM(offset_ + sizeof(OatQuickMethodHeader),
GetInstructionSetAlignment(compiled_method->GetInstructionSet()));
return offset_ + sizeof(OatQuickMethodHeader) + thumb_offset;
}
@@ -1010,17 +1017,16 @@
ReportWriteFailure("relative call thunk", it);
return false;
}
- uint32_t aligned_offset = compiled_method->AlignCode(offset_);
- uint32_t aligned_code_delta = aligned_offset - offset_;
- if (aligned_code_delta != 0) {
- if (!writer_->WriteCodeAlignment(out, aligned_code_delta)) {
+ uint32_t alignment_size = CodeAlignmentSize(offset_, *compiled_method);
+ if (alignment_size != 0) {
+ if (!writer_->WriteCodeAlignment(out, alignment_size)) {
ReportWriteFailure("code alignment padding", it);
return false;
}
- offset_ += aligned_code_delta;
+ offset_ += alignment_size;
DCHECK_OFFSET_();
}
- DCHECK_ALIGNED_PARAM(offset_,
+ DCHECK_ALIGNED_PARAM(offset_ + sizeof(OatQuickMethodHeader),
GetInstructionSetAlignment(compiled_method->GetInstructionSet()));
DCHECK_EQ(method_offsets.code_offset_,
offset_ + sizeof(OatQuickMethodHeader) + compiled_method->CodeDelta())