summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--compiler/linker/arm/relative_patcher_thumb2_test.cc33
-rw-r--r--compiler/linker/arm64/relative_patcher_arm64_test.cc70
-rw-r--r--compiler/linker/mips/relative_patcher_mips32r6_test.cc19
-rw-r--r--compiler/linker/mips/relative_patcher_mips_test.cc27
-rw-r--r--compiler/linker/mips64/relative_patcher_mips64_test.cc19
-rw-r--r--compiler/linker/multi_oat_relative_patcher.h2
-rw-r--r--compiler/linker/multi_oat_relative_patcher_test.cc2
-rw-r--r--compiler/linker/relative_patcher_test.h10
-rw-r--r--compiler/linker/x86/relative_patcher_x86_test.cc13
-rw-r--r--compiler/linker/x86_64/relative_patcher_x86_64_test.cc13
-rw-r--r--compiler/optimizing/loop_optimization.cc76
-rw-r--r--runtime/gc/collector/concurrent_copying.cc5
-rw-r--r--runtime/gc/verification.cc2
-rw-r--r--test/646-checker-hadd-short/src/Main.java129
14 files changed, 301 insertions, 119 deletions
diff --git a/compiler/linker/arm/relative_patcher_thumb2_test.cc b/compiler/linker/arm/relative_patcher_thumb2_test.cc
index af5fa40dc1..52e27afcf5 100644
--- a/compiler/linker/arm/relative_patcher_thumb2_test.cc
+++ b/compiler/linker/arm/relative_patcher_thumb2_test.cc
@@ -222,7 +222,7 @@ class Thumb2RelativePatcherTest : public RelativePatcherTest {
return result;
}
- void TestDexCacheReference(uint32_t dex_cache_arrays_begin, uint32_t element_offset);
+ void TestStringBssEntry(uint32_t bss_begin, uint32_t string_entry_offset);
void TestStringReference(uint32_t string_offset);
void CheckPcRelativePatch(const ArrayRef<const LinkerPatch>& patches, uint32_t target_offset);
@@ -290,15 +290,16 @@ const ArrayRef<const uint8_t> Thumb2RelativePatcherTest::kUnpatchedPcRelativeCod
kUnpatchedPcRelativeRawCode);
const uint32_t Thumb2RelativePatcherTest::kPcInsnOffset = 8u;
-void Thumb2RelativePatcherTest::TestDexCacheReference(uint32_t dex_cache_arrays_begin,
- uint32_t element_offset) {
- dex_cache_arrays_begin_ = dex_cache_arrays_begin;
+void Thumb2RelativePatcherTest::TestStringBssEntry(uint32_t bss_begin,
+ uint32_t string_entry_offset) {
+ constexpr uint32_t kStringIndex = 1u;
+ string_index_to_offset_map_.Put(kStringIndex, string_entry_offset);
+ bss_begin_ = bss_begin;
const LinkerPatch patches[] = {
- LinkerPatch::DexCacheArrayPatch(0u, nullptr, kPcInsnOffset, element_offset),
- LinkerPatch::DexCacheArrayPatch(4u, nullptr, kPcInsnOffset, element_offset),
+ LinkerPatch::StringBssEntryPatch(0u, nullptr, kPcInsnOffset, kStringIndex),
+ LinkerPatch::StringBssEntryPatch(4u, nullptr, kPcInsnOffset, kStringIndex),
};
- CheckPcRelativePatch(ArrayRef<const LinkerPatch>(patches),
- dex_cache_arrays_begin_ + element_offset);
+ CheckPcRelativePatch(ArrayRef<const LinkerPatch>(patches), bss_begin_ + string_entry_offset);
}
void Thumb2RelativePatcherTest::TestStringReference(uint32_t string_offset) {
@@ -534,23 +535,23 @@ TEST_F(Thumb2RelativePatcherTest, CallOtherJustTooFarBefore) {
EXPECT_TRUE(CheckThunk(thunk_offset));
}
-TEST_F(Thumb2RelativePatcherTest, DexCacheReference1) {
- TestDexCacheReference(0x00ff0000u, 0x00fcu);
+TEST_F(Thumb2RelativePatcherTest, StringBssEntry1) {
+ TestStringBssEntry(0x00ff0000u, 0x00fcu);
ASSERT_LT(GetMethodOffset(1u), 0xfcu);
}
-TEST_F(Thumb2RelativePatcherTest, DexCacheReference2) {
- TestDexCacheReference(0x02ff0000u, 0x05fcu);
+TEST_F(Thumb2RelativePatcherTest, StringBssEntry2) {
+ TestStringBssEntry(0x02ff0000u, 0x05fcu);
ASSERT_LT(GetMethodOffset(1u), 0xfcu);
}
-TEST_F(Thumb2RelativePatcherTest, DexCacheReference3) {
- TestDexCacheReference(0x08ff0000u, 0x08fcu);
+TEST_F(Thumb2RelativePatcherTest, StringBssEntry3) {
+ TestStringBssEntry(0x08ff0000u, 0x08fcu);
ASSERT_LT(GetMethodOffset(1u), 0xfcu);
}
-TEST_F(Thumb2RelativePatcherTest, DexCacheReference4) {
- TestDexCacheReference(0xd0ff0000u, 0x60fcu);
+TEST_F(Thumb2RelativePatcherTest, StringBssEntry4) {
+ TestStringBssEntry(0xd0ff0000u, 0x60fcu);
ASSERT_LT(GetMethodOffset(1u), 0xfcu);
}
diff --git a/compiler/linker/arm64/relative_patcher_arm64_test.cc b/compiler/linker/arm64/relative_patcher_arm64_test.cc
index b6549eefb3..5d02d449fe 100644
--- a/compiler/linker/arm64/relative_patcher_arm64_test.cc
+++ b/compiler/linker/arm64/relative_patcher_arm64_test.cc
@@ -249,12 +249,14 @@ class Arm64RelativePatcherTest : public RelativePatcherTest {
return GenNopsAndAdrpAndUse(num_nops, method_offset, target_offset, kLdrWInsn);
}
- void TestNopsAdrpLdr(size_t num_nops, uint32_t dex_cache_arrays_begin, uint32_t element_offset) {
- dex_cache_arrays_begin_ = dex_cache_arrays_begin;
+ void TestNopsAdrpLdr(size_t num_nops, uint32_t bss_begin, uint32_t string_entry_offset) {
+ constexpr uint32_t kStringIndex = 1u;
+ string_index_to_offset_map_.Put(kStringIndex, string_entry_offset);
+ bss_begin_ = bss_begin;
auto code = GenNopsAndAdrpLdr(num_nops, 0u, 0u); // Unpatched.
const LinkerPatch patches[] = {
- LinkerPatch::DexCacheArrayPatch(num_nops * 4u , nullptr, num_nops * 4u, element_offset),
- LinkerPatch::DexCacheArrayPatch(num_nops * 4u + 4u, nullptr, num_nops * 4u, element_offset),
+ LinkerPatch::StringBssEntryPatch(num_nops * 4u , nullptr, num_nops * 4u, kStringIndex),
+ LinkerPatch::StringBssEntryPatch(num_nops * 4u + 4u, nullptr, num_nops * 4u, kStringIndex),
};
AddCompiledMethod(MethodRef(1u),
ArrayRef<const uint8_t>(code),
@@ -262,7 +264,7 @@ class Arm64RelativePatcherTest : public RelativePatcherTest {
Link();
uint32_t method1_offset = GetMethodOffset(1u);
- uint32_t target_offset = dex_cache_arrays_begin_ + element_offset;
+ uint32_t target_offset = bss_begin_ + string_entry_offset;
auto expected_code = GenNopsAndAdrpLdr(num_nops, method1_offset, target_offset);
EXPECT_TRUE(CheckLinkedMethod(MethodRef(1u), ArrayRef<const uint8_t>(expected_code)));
}
@@ -293,14 +295,16 @@ class Arm64RelativePatcherTest : public RelativePatcherTest {
void PrepareNopsAdrpInsn2Ldr(size_t num_nops,
uint32_t insn2,
- uint32_t dex_cache_arrays_begin,
- uint32_t element_offset) {
- dex_cache_arrays_begin_ = dex_cache_arrays_begin;
+ uint32_t bss_begin,
+ uint32_t string_entry_offset) {
+ constexpr uint32_t kStringIndex = 1u;
+ string_index_to_offset_map_.Put(kStringIndex, string_entry_offset);
+ bss_begin_ = bss_begin;
auto code = GenNopsAndAdrpLdr(num_nops, 0u, 0u); // Unpatched.
InsertInsn(&code, num_nops * 4u + 4u, insn2);
const LinkerPatch patches[] = {
- LinkerPatch::DexCacheArrayPatch(num_nops * 4u , nullptr, num_nops * 4u, element_offset),
- LinkerPatch::DexCacheArrayPatch(num_nops * 4u + 8u, nullptr, num_nops * 4u, element_offset),
+ LinkerPatch::StringBssEntryPatch(num_nops * 4u , nullptr, num_nops * 4u, kStringIndex),
+ LinkerPatch::StringBssEntryPatch(num_nops * 4u + 8u, nullptr, num_nops * 4u, kStringIndex),
};
AddCompiledMethod(MethodRef(1u),
ArrayRef<const uint8_t>(code),
@@ -376,15 +380,15 @@ class Arm64RelativePatcherTest : public RelativePatcherTest {
void TestAdrpInsn2Ldr(uint32_t insn2,
uint32_t adrp_offset,
bool has_thunk,
- uint32_t dex_cache_arrays_begin,
- uint32_t element_offset) {
+ uint32_t bss_begin,
+ uint32_t string_entry_offset) {
uint32_t method1_offset =
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;
- PrepareNopsAdrpInsn2Ldr(num_nops, insn2, dex_cache_arrays_begin, element_offset);
- uint32_t target_offset = dex_cache_arrays_begin_ + element_offset;
+ PrepareNopsAdrpInsn2Ldr(num_nops, insn2, bss_begin, string_entry_offset);
+ uint32_t target_offset = bss_begin_ + string_entry_offset;
if (has_thunk) {
TestNopsAdrpInsn2AndUseHasThunk(num_nops, insn2, target_offset, kLdrWInsn);
} else {
@@ -395,33 +399,33 @@ class Arm64RelativePatcherTest : public RelativePatcherTest {
void TestAdrpLdurLdr(uint32_t adrp_offset,
bool has_thunk,
- uint32_t dex_cache_arrays_begin,
- uint32_t element_offset) {
- TestAdrpInsn2Ldr(kLdurInsn, adrp_offset, has_thunk, dex_cache_arrays_begin, element_offset);
+ uint32_t bss_begin,
+ uint32_t string_entry_offset) {
+ TestAdrpInsn2Ldr(kLdurInsn, adrp_offset, has_thunk, bss_begin, string_entry_offset);
}
void TestAdrpLdrPcRelLdr(uint32_t pcrel_ldr_insn,
int32_t pcrel_disp,
uint32_t adrp_offset,
bool has_thunk,
- uint32_t dex_cache_arrays_begin,
- uint32_t element_offset) {
+ uint32_t bss_begin,
+ uint32_t string_entry_offset) {
ASSERT_LT(pcrel_disp, 0x100000);
ASSERT_GE(pcrel_disp, -0x100000);
ASSERT_EQ(pcrel_disp & 0x3, 0);
uint32_t insn2 = pcrel_ldr_insn | (((static_cast<uint32_t>(pcrel_disp) >> 2) & 0x7ffffu) << 5);
- TestAdrpInsn2Ldr(insn2, adrp_offset, has_thunk, dex_cache_arrays_begin, element_offset);
+ TestAdrpInsn2Ldr(insn2, adrp_offset, has_thunk, bss_begin, string_entry_offset);
}
void TestAdrpLdrSpRelLdr(uint32_t sprel_ldr_insn,
uint32_t sprel_disp_in_load_units,
uint32_t adrp_offset,
bool has_thunk,
- uint32_t dex_cache_arrays_begin,
- uint32_t element_offset) {
+ uint32_t bss_begin,
+ uint32_t string_entry_offset) {
ASSERT_LT(sprel_disp_in_load_units, 0x1000u);
uint32_t insn2 = sprel_ldr_insn | ((sprel_disp_in_load_units & 0xfffu) << 10);
- TestAdrpInsn2Ldr(insn2, adrp_offset, has_thunk, dex_cache_arrays_begin, element_offset);
+ TestAdrpInsn2Ldr(insn2, adrp_offset, has_thunk, bss_begin, string_entry_offset);
}
void TestAdrpInsn2Add(uint32_t insn2,
@@ -726,19 +730,19 @@ TEST_F(Arm64RelativePatcherTestDefault, CallOtherJustTooFarBefore) {
EXPECT_TRUE(CheckThunk(thunk_offset));
}
-TEST_F(Arm64RelativePatcherTestDefault, DexCacheReference1) {
+TEST_F(Arm64RelativePatcherTestDefault, StringBssEntry1) {
TestNopsAdrpLdr(0u, 0x12345678u, 0x1234u);
}
-TEST_F(Arm64RelativePatcherTestDefault, DexCacheReference2) {
+TEST_F(Arm64RelativePatcherTestDefault, StringBssEntry2) {
TestNopsAdrpLdr(0u, -0x12345678u, 0x4444u);
}
-TEST_F(Arm64RelativePatcherTestDefault, DexCacheReference3) {
+TEST_F(Arm64RelativePatcherTestDefault, StringBssEntry3) {
TestNopsAdrpLdr(0u, 0x12345000u, 0x3ffcu);
}
-TEST_F(Arm64RelativePatcherTestDefault, DexCacheReference4) {
+TEST_F(Arm64RelativePatcherTestDefault, StringBssEntry4) {
TestNopsAdrpLdr(0u, 0x12345000u, 0x4000u);
}
@@ -763,7 +767,7 @@ TEST_F(Arm64RelativePatcherTestDefault, StringReference4) {
test(0xff4u, disp2) test(0xff8u, disp2) test(0xffcu, disp2) test(0x1000u, disp2)
#define DEFAULT_LDUR_LDR_TEST(adrp_offset, disp) \
- TEST_F(Arm64RelativePatcherTestDefault, DexCacheReference ## adrp_offset ## Ldur ## disp) { \
+ TEST_F(Arm64RelativePatcherTestDefault, StringBssEntry ## adrp_offset ## Ldur ## disp) { \
bool has_thunk = ((adrp_offset) == 0xff8u || (adrp_offset) == 0xffcu); \
TestAdrpLdurLdr(adrp_offset, has_thunk, 0x12345678u, disp); \
}
@@ -771,7 +775,7 @@ TEST_F(Arm64RelativePatcherTestDefault, StringReference4) {
TEST_FOR_OFFSETS(DEFAULT_LDUR_LDR_TEST, 0x1234, 0x1238)
#define DENVER64_LDUR_LDR_TEST(adrp_offset, disp) \
- TEST_F(Arm64RelativePatcherTestDenver64, DexCacheReference ## adrp_offset ## Ldur ## disp) { \
+ TEST_F(Arm64RelativePatcherTestDenver64, StringBssEntry ## adrp_offset ## Ldur ## disp) { \
TestAdrpLdurLdr(adrp_offset, false, 0x12345678u, disp); \
}
@@ -779,7 +783,7 @@ TEST_FOR_OFFSETS(DENVER64_LDUR_LDR_TEST, 0x1234, 0x1238)
// LDR <Wt>, <label> is always aligned. We should never have to use a fixup.
#define LDRW_PCREL_LDR_TEST(adrp_offset, disp) \
- TEST_F(Arm64RelativePatcherTestDefault, DexCacheReference ## adrp_offset ## WPcRel ## disp) { \
+ TEST_F(Arm64RelativePatcherTestDefault, StringBssEntry ## adrp_offset ## WPcRel ## disp) { \
TestAdrpLdrPcRelLdr(kLdrWPcRelInsn, disp, adrp_offset, false, 0x12345678u, 0x1234u); \
}
@@ -787,7 +791,7 @@ TEST_FOR_OFFSETS(LDRW_PCREL_LDR_TEST, 0x1234, 0x1238)
// LDR <Xt>, <label> is aligned when offset + displacement is a multiple of 8.
#define LDRX_PCREL_LDR_TEST(adrp_offset, disp) \
- TEST_F(Arm64RelativePatcherTestDefault, DexCacheReference ## adrp_offset ## XPcRel ## disp) { \
+ TEST_F(Arm64RelativePatcherTestDefault, StringBssEntry ## adrp_offset ## XPcRel ## disp) { \
bool unaligned = !IsAligned<8u>((adrp_offset) + 4u + static_cast<uint32_t>(disp)); \
bool has_thunk = ((adrp_offset) == 0xff8u || (adrp_offset) == 0xffcu) && unaligned; \
TestAdrpLdrPcRelLdr(kLdrXPcRelInsn, disp, adrp_offset, has_thunk, 0x12345678u, 0x1234u); \
@@ -797,14 +801,14 @@ TEST_FOR_OFFSETS(LDRX_PCREL_LDR_TEST, 0x1234, 0x1238)
// LDR <Wt>, [SP, #<pimm>] and LDR <Xt>, [SP, #<pimm>] are always aligned. No fixup needed.
#define LDRW_SPREL_LDR_TEST(adrp_offset, disp) \
- TEST_F(Arm64RelativePatcherTestDefault, DexCacheReference ## adrp_offset ## WSpRel ## disp) { \
+ TEST_F(Arm64RelativePatcherTestDefault, StringBssEntry ## adrp_offset ## WSpRel ## disp) { \
TestAdrpLdrSpRelLdr(kLdrWSpRelInsn, (disp) >> 2, adrp_offset, false, 0x12345678u, 0x1234u); \
}
TEST_FOR_OFFSETS(LDRW_SPREL_LDR_TEST, 0, 4)
#define LDRX_SPREL_LDR_TEST(adrp_offset, disp) \
- TEST_F(Arm64RelativePatcherTestDefault, DexCacheReference ## adrp_offset ## XSpRel ## disp) { \
+ TEST_F(Arm64RelativePatcherTestDefault, StringBssEntry ## adrp_offset ## XSpRel ## disp) { \
TestAdrpLdrSpRelLdr(kLdrXSpRelInsn, (disp) >> 3, adrp_offset, false, 0x12345678u, 0x1234u); \
}
diff --git a/compiler/linker/mips/relative_patcher_mips32r6_test.cc b/compiler/linker/mips/relative_patcher_mips32r6_test.cc
index 474eb73e08..63ad8a58d5 100644
--- a/compiler/linker/mips/relative_patcher_mips32r6_test.cc
+++ b/compiler/linker/mips/relative_patcher_mips32r6_test.cc
@@ -37,7 +37,7 @@ class Mips32r6RelativePatcherTest : public RelativePatcherTest {
}
void CheckPcRelativePatch(const ArrayRef<const LinkerPatch>& patches, uint32_t target_offset);
- void TestDexCacheReference(uint32_t dex_cache_arrays_begin, uint32_t element_offset);
+ void TestStringBssEntry(uint32_t bss_begin, uint32_t string_entry_offset);
void TestStringReference(uint32_t string_offset);
};
@@ -69,14 +69,15 @@ void Mips32r6RelativePatcherTest::CheckPcRelativePatch(const ArrayRef<const Link
EXPECT_TRUE(CheckLinkedMethod(MethodRef(1u), ArrayRef<const uint8_t>(expected_code)));
}
-void Mips32r6RelativePatcherTest::TestDexCacheReference(uint32_t dex_cache_arrays_begin,
- uint32_t element_offset) {
- dex_cache_arrays_begin_ = dex_cache_arrays_begin;
+void Mips32r6RelativePatcherTest::TestStringBssEntry(uint32_t bss_begin,
+ uint32_t string_entry_offset) {
+ constexpr uint32_t kStringIndex = 1u;
+ string_index_to_offset_map_.Put(kStringIndex, string_entry_offset);
+ bss_begin_ = bss_begin;
LinkerPatch patches[] = {
- LinkerPatch::DexCacheArrayPatch(kLiteralOffset, nullptr, kAnchorOffset, element_offset)
+ LinkerPatch::StringBssEntryPatch(kLiteralOffset, nullptr, kAnchorOffset, kStringIndex)
};
- CheckPcRelativePatch(ArrayRef<const LinkerPatch>(patches),
- dex_cache_arrays_begin_ + element_offset);
+ CheckPcRelativePatch(ArrayRef<const LinkerPatch>(patches), bss_begin_ + string_entry_offset);
}
void Mips32r6RelativePatcherTest::TestStringReference(uint32_t string_offset) {
@@ -88,8 +89,8 @@ void Mips32r6RelativePatcherTest::TestStringReference(uint32_t string_offset) {
CheckPcRelativePatch(ArrayRef<const LinkerPatch>(patches), string_offset);
}
-TEST_F(Mips32r6RelativePatcherTest, DexCacheReference) {
- TestDexCacheReference(/* dex_cache_arrays_begin */ 0x12345678, /* element_offset */ 0x1234);
+TEST_F(Mips32r6RelativePatcherTest, StringBssEntry) {
+ TestStringBssEntry(/* bss_begin */ 0x12345678, /* string_entry_offset */ 0x1234);
}
TEST_F(Mips32r6RelativePatcherTest, StringReference) {
diff --git a/compiler/linker/mips/relative_patcher_mips_test.cc b/compiler/linker/mips/relative_patcher_mips_test.cc
index b0d1294cf4..961b31266f 100644
--- a/compiler/linker/mips/relative_patcher_mips_test.cc
+++ b/compiler/linker/mips/relative_patcher_mips_test.cc
@@ -20,10 +20,6 @@
namespace art {
namespace linker {
-// We'll maximize the range of a single load instruction for dex cache array accesses
-// by aligning offset -32768 with the offset of the first used element.
-static constexpr uint32_t kDexCacheArrayLwOffset = 0x8000;
-
class MipsRelativePatcherTest : public RelativePatcherTest {
public:
MipsRelativePatcherTest() : RelativePatcherTest(kMips, "mips32r2") {}
@@ -41,7 +37,7 @@ class MipsRelativePatcherTest : public RelativePatcherTest {
}
void CheckPcRelativePatch(const ArrayRef<const LinkerPatch>& patches, uint32_t target_offset);
- void TestDexCacheReference(uint32_t dex_cache_arrays_begin, uint32_t element_offset);
+ void TestStringBssEntry(uint32_t bss_begin, uint32_t string_entry_offset);
void TestStringReference(uint32_t string_offset);
};
@@ -65,9 +61,7 @@ void MipsRelativePatcherTest::CheckPcRelativePatch(const ArrayRef<const LinkerPa
ASSERT_TRUE(result.first);
uint32_t diff = target_offset - (result.second + kAnchorOffset);
- if (patches[0].GetType() == LinkerPatch::Type::kDexCacheArray) {
- diff += kDexCacheArrayLwOffset;
- }
+ CHECK_NE(patches[0].GetType(), LinkerPatch::Type::kDexCacheArray);
diff += (diff & 0x8000) << 1; // Account for sign extension in addiu.
const uint8_t expected_code[] = {
@@ -79,14 +73,15 @@ void MipsRelativePatcherTest::CheckPcRelativePatch(const ArrayRef<const LinkerPa
EXPECT_TRUE(CheckLinkedMethod(MethodRef(1u), ArrayRef<const uint8_t>(expected_code)));
}
-void MipsRelativePatcherTest::TestDexCacheReference(uint32_t dex_cache_arrays_begin,
- uint32_t element_offset) {
- dex_cache_arrays_begin_ = dex_cache_arrays_begin;
+void MipsRelativePatcherTest::TestStringBssEntry(uint32_t bss_begin,
+ uint32_t string_entry_offset) {
+ constexpr uint32_t kStringIndex = 1u;
+ string_index_to_offset_map_.Put(kStringIndex, string_entry_offset);
+ bss_begin_ = bss_begin;
LinkerPatch patches[] = {
- LinkerPatch::DexCacheArrayPatch(kLiteralOffset, nullptr, kAnchorOffset, element_offset)
+ LinkerPatch::StringBssEntryPatch(kLiteralOffset, nullptr, kAnchorOffset, kStringIndex)
};
- CheckPcRelativePatch(ArrayRef<const LinkerPatch>(patches),
- dex_cache_arrays_begin_ + element_offset);
+ CheckPcRelativePatch(ArrayRef<const LinkerPatch>(patches), bss_begin_ + string_entry_offset);
}
void MipsRelativePatcherTest::TestStringReference(uint32_t string_offset) {
@@ -98,8 +93,8 @@ void MipsRelativePatcherTest::TestStringReference(uint32_t string_offset) {
CheckPcRelativePatch(ArrayRef<const LinkerPatch>(patches), string_offset);
}
-TEST_F(MipsRelativePatcherTest, DexCacheReference) {
- TestDexCacheReference(/* dex_cache_arrays_begin */ 0x12345678, /* element_offset */ 0x1234);
+TEST_F(MipsRelativePatcherTest, StringBssEntry) {
+ TestStringBssEntry(/* bss_begin */ 0x12345678, /* string_entry_offset */ 0x1234);
}
TEST_F(MipsRelativePatcherTest, StringReference) {
diff --git a/compiler/linker/mips64/relative_patcher_mips64_test.cc b/compiler/linker/mips64/relative_patcher_mips64_test.cc
index c3170584e4..9c9e24a96e 100644
--- a/compiler/linker/mips64/relative_patcher_mips64_test.cc
+++ b/compiler/linker/mips64/relative_patcher_mips64_test.cc
@@ -39,7 +39,7 @@ class Mips64RelativePatcherTest : public RelativePatcherTest {
}
void CheckPcRelativePatch(const ArrayRef<const LinkerPatch>& patches, uint32_t target_offset);
- void TestDexCacheReference(uint32_t dex_cache_arrays_begin, uint32_t element_offset);
+ void TestStringBssEntry(uint32_t bss_begin, uint32_t string_entry_offset);
void TestStringReference(uint32_t string_offset);
};
@@ -76,18 +76,19 @@ void Mips64RelativePatcherTest::CheckPcRelativePatch(const ArrayRef<const Linker
EXPECT_TRUE(CheckLinkedMethod(MethodRef(1u), ArrayRef<const uint8_t>(expected_code)));
}
-void Mips64RelativePatcherTest::TestDexCacheReference(uint32_t dex_cache_arrays_begin,
- uint32_t element_offset) {
- dex_cache_arrays_begin_ = dex_cache_arrays_begin;
+void Mips64RelativePatcherTest::TestStringBssEntry(uint32_t bss_begin,
+ uint32_t string_entry_offset) {
+ constexpr uint32_t kStringIndex = 1u;
+ string_index_to_offset_map_.Put(kStringIndex, string_entry_offset);
+ bss_begin_ = bss_begin;
LinkerPatch patches[] = {
- LinkerPatch::DexCacheArrayPatch(kLiteralOffset, nullptr, kAnchorOffset, element_offset)
+ LinkerPatch::StringBssEntryPatch(kLiteralOffset, nullptr, kAnchorOffset, kStringIndex)
};
- CheckPcRelativePatch(ArrayRef<const LinkerPatch>(patches),
- dex_cache_arrays_begin_ + element_offset);
+ CheckPcRelativePatch(ArrayRef<const LinkerPatch>(patches), bss_begin_ + string_entry_offset);
}
-TEST_F(Mips64RelativePatcherTest, DexCacheReference) {
- TestDexCacheReference(/* dex_cache_arrays_begin */ 0x12345678, /* element_offset */ 0x1234);
+TEST_F(Mips64RelativePatcherTest, StringBssEntry) {
+ TestStringBssEntry(/* bss_begin */ 0x12345678, /* string_entry_offset */ 0x1234);
}
TEST_F(Mips64RelativePatcherTest, CallOther) {
diff --git a/compiler/linker/multi_oat_relative_patcher.h b/compiler/linker/multi_oat_relative_patcher.h
index 247b29017e..bdc1ee1d27 100644
--- a/compiler/linker/multi_oat_relative_patcher.h
+++ b/compiler/linker/multi_oat_relative_patcher.h
@@ -102,7 +102,7 @@ class MultiOatRelativePatcher FINAL {
relative_patcher_->PatchCall(code, literal_offset, patch_offset, target_offset);
}
- // Wrapper around RelativePatcher::PatchDexCacheReference(), doing offset adjustment.
+ // Wrapper around RelativePatcher::PatchPcRelativeReference(), doing offset adjustment.
void PatchPcRelativeReference(std::vector<uint8_t>* code,
const LinkerPatch& patch,
uint32_t patch_offset,
diff --git a/compiler/linker/multi_oat_relative_patcher_test.cc b/compiler/linker/multi_oat_relative_patcher_test.cc
index 951588a857..615b2b97be 100644
--- a/compiler/linker/multi_oat_relative_patcher_test.cc
+++ b/compiler/linker/multi_oat_relative_patcher_test.cc
@@ -282,7 +282,7 @@ TEST_F(MultiOatRelativePatcherTest, Patch) {
uint32_t method2_patch_offset = 0x7654u;
uint32_t method2_target_offset = 0xccccu;
LinkerPatch method2_patch =
- LinkerPatch::DexCacheArrayPatch(method2_literal_offset, nullptr, 0u, 1234u);
+ LinkerPatch::StringBssEntryPatch(method2_literal_offset, nullptr, 0u, 1u);
patcher_.PatchPcRelativeReference(
&code, method2_patch, method2_patch_offset, method2_target_offset);
DCHECK_EQ(method2_literal_offset, mock_->last_literal_offset_);
diff --git a/compiler/linker/relative_patcher_test.h b/compiler/linker/relative_patcher_test.h
index d9a87a0cfd..bff68080c6 100644
--- a/compiler/linker/relative_patcher_test.h
+++ b/compiler/linker/relative_patcher_test.h
@@ -31,6 +31,7 @@
#include "method_reference.h"
#include "oat.h"
#include "oat_quick_method_header.h"
+#include "string_reference.h"
#include "vector_output_stream.h"
namespace art {
@@ -61,7 +62,7 @@ class RelativePatcherTest : public testing::Test {
features_(InstructionSetFeatures::FromVariant(instruction_set, variant, &error_msg_)),
method_offset_map_(),
patcher_(RelativePatcher::Create(instruction_set, features_.get(), &method_offset_map_)),
- dex_cache_arrays_begin_(0u),
+ bss_begin_(0u),
compiled_method_refs_(),
compiled_methods_(),
patched_code_(),
@@ -157,8 +158,9 @@ class RelativePatcherTest : public testing::Test {
result.first ? result.second : kTrampolineOffset + compiled_method->CodeDelta();
patcher_->PatchCall(&patched_code_, patch.LiteralOffset(),
offset + patch.LiteralOffset(), target_offset);
- } else if (patch.GetType() == LinkerPatch::Type::kDexCacheArray) {
- uint32_t target_offset = dex_cache_arrays_begin_ + patch.TargetDexCacheElementOffset();
+ } else if (patch.GetType() == LinkerPatch::Type::kStringBssEntry) {
+ uint32_t target_offset =
+ bss_begin_ + string_index_to_offset_map_.Get(patch.TargetStringIndex().index_);
patcher_->PatchPcRelativeReference(&patched_code_,
patch,
offset + patch.LiteralOffset(),
@@ -276,7 +278,7 @@ class RelativePatcherTest : public testing::Test {
std::unique_ptr<const InstructionSetFeatures> features_;
MethodOffsetMap method_offset_map_;
std::unique_ptr<RelativePatcher> patcher_;
- uint32_t dex_cache_arrays_begin_;
+ uint32_t bss_begin_;
SafeMap<uint32_t, uint32_t> string_index_to_offset_map_;
std::vector<MethodReference> compiled_method_refs_;
std::vector<std::unique_ptr<CompiledMethod>> compiled_methods_;
diff --git a/compiler/linker/x86/relative_patcher_x86_test.cc b/compiler/linker/x86/relative_patcher_x86_test.cc
index 2a44b7990e..0bd9de8e15 100644
--- a/compiler/linker/x86/relative_patcher_x86_test.cc
+++ b/compiler/linker/x86/relative_patcher_x86_test.cc
@@ -107,9 +107,11 @@ TEST_F(X86RelativePatcherTest, CallTrampoline) {
EXPECT_TRUE(CheckLinkedMethod(MethodRef(1u), ArrayRef<const uint8_t>(expected_code)));
}
-TEST_F(X86RelativePatcherTest, DexCacheReference) {
- dex_cache_arrays_begin_ = 0x12345678;
- constexpr size_t kElementOffset = 0x1234;
+TEST_F(X86RelativePatcherTest, StringBssEntry) {
+ bss_begin_ = 0x12345678;
+ constexpr size_t kStringEntryOffset = 0x1234;
+ constexpr uint32_t kStringIndex = 1u;
+ string_index_to_offset_map_.Put(kStringIndex, kStringEntryOffset);
static const uint8_t raw_code[] = {
0xe8, 0x00, 0x00, 0x00, 0x00, // call +0
0x5b, // pop ebx
@@ -118,15 +120,14 @@ TEST_F(X86RelativePatcherTest, DexCacheReference) {
constexpr uint32_t anchor_offset = 5u; // After call +0.
ArrayRef<const uint8_t> code(raw_code);
LinkerPatch patches[] = {
- LinkerPatch::DexCacheArrayPatch(code.size() - 4u, nullptr, anchor_offset, kElementOffset),
+ LinkerPatch::StringBssEntryPatch(code.size() - 4u, nullptr, anchor_offset, kStringIndex),
};
AddCompiledMethod(MethodRef(1u), code, ArrayRef<const LinkerPatch>(patches));
Link();
auto result = method_offset_map_.FindMethodOffset(MethodRef(1u));
ASSERT_TRUE(result.first);
- uint32_t diff =
- dex_cache_arrays_begin_ + kElementOffset - (result.second + anchor_offset);
+ uint32_t diff = bss_begin_ + kStringEntryOffset - (result.second + anchor_offset);
static const uint8_t expected_code[] = {
0xe8, 0x00, 0x00, 0x00, 0x00, // call +0
0x5b, // pop ebx
diff --git a/compiler/linker/x86_64/relative_patcher_x86_64_test.cc b/compiler/linker/x86_64/relative_patcher_x86_64_test.cc
index 2b46453255..6d6bb40fb4 100644
--- a/compiler/linker/x86_64/relative_patcher_x86_64_test.cc
+++ b/compiler/linker/x86_64/relative_patcher_x86_64_test.cc
@@ -127,19 +127,20 @@ TEST_F(X86_64RelativePatcherTest, CallTrampoline) {
EXPECT_TRUE(CheckLinkedMethod(MethodRef(1u), ArrayRef<const uint8_t>(expected_code)));
}
-TEST_F(X86_64RelativePatcherTest, DexCacheReference) {
- dex_cache_arrays_begin_ = 0x12345678;
- constexpr size_t kElementOffset = 0x1234;
+TEST_F(X86_64RelativePatcherTest, StringBssEntry) {
+ bss_begin_ = 0x12345678;
+ constexpr size_t kStringEntryOffset = 0x1234;
+ constexpr uint32_t kStringIndex = 1u;
+ string_index_to_offset_map_.Put(kStringIndex, kStringEntryOffset);
LinkerPatch patches[] = {
- LinkerPatch::DexCacheArrayPatch(kDexCacheLoadCode.size() - 4u, nullptr, 0u, kElementOffset),
+ LinkerPatch::StringBssEntryPatch(kDexCacheLoadCode.size() - 4u, nullptr, 0u, kStringIndex),
};
AddCompiledMethod(MethodRef(1u), kDexCacheLoadCode, ArrayRef<const LinkerPatch>(patches));
Link();
auto result = method_offset_map_.FindMethodOffset(MethodRef(1u));
ASSERT_TRUE(result.first);
- uint32_t diff =
- dex_cache_arrays_begin_ + kElementOffset - (result.second + kDexCacheLoadCode.size());
+ uint32_t diff = bss_begin_ + kStringEntryOffset - (result.second + kDexCacheLoadCode.size());
static const uint8_t expected_code[] = {
0x8b, 0x05,
static_cast<uint8_t>(diff),
diff --git a/compiler/optimizing/loop_optimization.cc b/compiler/optimizing/loop_optimization.cc
index 4067aa3468..963df5a938 100644
--- a/compiler/optimizing/loop_optimization.cc
+++ b/compiler/optimizing/loop_optimization.cc
@@ -173,6 +173,51 @@ static bool IsZeroExtensionAndGet(HInstruction* instruction,
return false;
}
+// Detect up to two instructions a and b, and an acccumulated constant c.
+static bool IsAddConstHelper(HInstruction* instruction,
+ /*out*/ HInstruction** a,
+ /*out*/ HInstruction** b,
+ /*out*/ int64_t* c,
+ int32_t depth) {
+ static constexpr int32_t kMaxDepth = 8; // don't search too deep
+ int64_t value = 0;
+ if (IsInt64AndGet(instruction, &value)) {
+ *c += value;
+ return true;
+ } else if (instruction->IsAdd() && depth <= kMaxDepth) {
+ return IsAddConstHelper(instruction->InputAt(0), a, b, c, depth + 1) &&
+ IsAddConstHelper(instruction->InputAt(1), a, b, c, depth + 1);
+ } else if (*a == nullptr) {
+ *a = instruction;
+ return true;
+ } else if (*b == nullptr) {
+ *b = instruction;
+ return true;
+ }
+ return false; // too many non-const operands
+}
+
+// Detect a + b + c for an optional constant c.
+static bool IsAddConst(HInstruction* instruction,
+ /*out*/ HInstruction** a,
+ /*out*/ HInstruction** b,
+ /*out*/ int64_t* c) {
+ if (instruction->IsAdd()) {
+ // Try to find a + b and accumulated c.
+ if (IsAddConstHelper(instruction->InputAt(0), a, b, c, /*depth*/ 0) &&
+ IsAddConstHelper(instruction->InputAt(1), a, b, c, /*depth*/ 0) &&
+ *b != nullptr) {
+ return true;
+ }
+ // Found a + b.
+ *a = instruction->InputAt(0);
+ *b = instruction->InputAt(1);
+ *c = 0;
+ return true;
+ }
+ return false;
+}
+
// Test vector restrictions.
static bool HasVectorRestrictions(uint64_t restrictions, uint64_t tested) {
return (restrictions & tested) != 0;
@@ -1215,24 +1260,23 @@ bool HLoopOptimization::VectorizeHalvingAddIdiom(LoopNode* node,
// Test for top level arithmetic shift right x >> 1 or logical shift right x >>> 1
// (note whether the sign bit in higher precision is shifted in has no effect
// on the narrow precision computed by the idiom).
- int64_t value = 0;
+ int64_t distance = 0;
if ((instruction->IsShr() ||
instruction->IsUShr()) &&
- IsInt64AndGet(instruction->InputAt(1), /*out*/ &value) && value == 1) {
- //
- // TODO: make following code less sensitive to associativity and commutativity differences.
- //
- HInstruction* x = instruction->InputAt(0);
- // Test for an optional rounding part (x + 1) >> 1.
- bool is_rounded = false;
- if (x->IsAdd() && IsInt64AndGet(x->InputAt(1), /*out*/ &value) && value == 1) {
- x = x->InputAt(0);
- is_rounded = true;
- }
- // Test for a core addition (a + b) >> 1 (possibly rounded), either unsigned or signed.
- if (x->IsAdd()) {
- HInstruction* a = x->InputAt(0);
- HInstruction* b = x->InputAt(1);
+ IsInt64AndGet(instruction->InputAt(1), /*out*/ &distance) && distance == 1) {
+ // Test for (a + b + c) >> 1 for optional constant c.
+ HInstruction* a = nullptr;
+ HInstruction* b = nullptr;
+ int64_t c = 0;
+ if (IsAddConst(instruction->InputAt(0), /*out*/ &a, /*out*/ &b, /*out*/ &c)) {
+ // Accept c == 1 (rounded) or c == 0 (not rounded).
+ bool is_rounded = false;
+ if (c == 1) {
+ is_rounded = true;
+ } else if (c != 0) {
+ return false;
+ }
+ // Accept consistent zero or sign extension on operands a and b.
HInstruction* r = nullptr;
HInstruction* s = nullptr;
bool is_unsigned = false;
diff --git a/runtime/gc/collector/concurrent_copying.cc b/runtime/gc/collector/concurrent_copying.cc
index b0218b51fa..ef843c6650 100644
--- a/runtime/gc/collector/concurrent_copying.cc
+++ b/runtime/gc/collector/concurrent_copying.cc
@@ -2070,7 +2070,10 @@ inline void ConcurrentCopying::Process(mirror::Object* obj, MemberOffset offset)
DCHECK_EQ(Thread::Current(), thread_running_gc_);
mirror::Object* ref = obj->GetFieldObject<
mirror::Object, kVerifyNone, kWithoutReadBarrier, false>(offset);
- mirror::Object* to_ref = Mark</*kGrayImmuneObject*/false, /*kFromGCThread*/true>(ref);
+ mirror::Object* to_ref = Mark</*kGrayImmuneObject*/false, /*kFromGCThread*/true>(
+ ref,
+ /*holder*/ obj,
+ offset);
if (to_ref == ref) {
return;
}
diff --git a/runtime/gc/verification.cc b/runtime/gc/verification.cc
index 7b31c8a259..c14f250528 100644
--- a/runtime/gc/verification.cc
+++ b/runtime/gc/verification.cc
@@ -86,8 +86,8 @@ void Verification::LogHeapCorruption(ObjPtr<mirror::Object> holder,
std::ostringstream oss;
oss << "GC tried to mark invalid reference " << ref << std::endl;
oss << DumpObjectInfo(ref, "ref") << "\n";
+ oss << DumpObjectInfo(holder.Ptr(), "holder");
if (holder != nullptr) {
- oss << DumpObjectInfo(holder.Ptr(), "holder");
mirror::Class* holder_klass = holder->GetClass<kVerifyNone, kWithoutReadBarrier>();
if (IsValidClass(holder_klass)) {
oss << "field_offset=" << offset.Uint32Value();
diff --git a/test/646-checker-hadd-short/src/Main.java b/test/646-checker-hadd-short/src/Main.java
index db495f6433..4e6b4bde8b 100644
--- a/test/646-checker-hadd-short/src/Main.java
+++ b/test/646-checker-hadd-short/src/Main.java
@@ -49,6 +49,34 @@ public class Main {
}
}
+ /// CHECK-START: void Main.halving_add_signed_alt(short[], short[], short[]) loop_optimization (before)
+ /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none
+ /// CHECK-DAG: <<I10:i\d+>> IntConstant 10 loop:none
+ /// CHECK-DAG: <<M10:i\d+>> IntConstant -10 loop:none
+ /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none
+ /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<I10>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Add2:i\d+>> Add [<<Get2>>,<<M10>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Add3:i\d+>> Add [<<Add1>>,<<Add2>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add3>>,<<I1>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none
+ //
+ /// CHECK-START-ARM64: void Main.halving_add_signed_alt(short[], short[], short[]) loop_optimization (after)
+ /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none
+ /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:false loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none
+ private static void halving_add_signed_alt(short[] b1, short[] b2, short[] bo) {
+ int min_length = Math.min(bo.length, Math.min(b1.length, b2.length));
+ for (int i = 0; i < min_length; i++) {
+ // Cancelling constant computations do not confuse recognition.
+ bo[i] = (short) (((b1[i] + 10) + (b2[i] - 10)) >> 1);
+ }
+ }
+
/// CHECK-START: void Main.halving_add_unsigned(short[], short[], short[]) loop_optimization (before)
/// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none
/// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none
@@ -99,6 +127,59 @@ public class Main {
}
}
+ /// CHECK-START: void Main.rounding_halving_add_signed_alt(short[], short[], short[]) loop_optimization (before)
+ /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none
+ /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none
+ /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<I1>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Add2:i\d+>> Add [<<Add1>>,<<Get2>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add2>>,<<I1>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none
+ //
+ /// CHECK-START-ARM64: void Main.rounding_halving_add_signed_alt(short[], short[], short[]) loop_optimization (after)
+ /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none
+ /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:true loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none
+ private static void rounding_halving_add_signed_alt(short[] b1, short[] b2, short[] bo) {
+ int min_length = Math.min(bo.length, Math.min(b1.length, b2.length));
+ for (int i = 0; i < min_length; i++) {
+ // Slightly different order in idiom does not confuse recognition.
+ bo[i] = (short) (((1 + b1[i]) + b2[i]) >> 1);
+ }
+ }
+
+ /// CHECK-START: void Main.rounding_halving_add_signed_alt2(short[], short[], short[]) loop_optimization (before)
+ /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none
+ /// CHECK-DAG: <<I10:i\d+>> IntConstant 10 loop:none
+ /// CHECK-DAG: <<M9:i\d+>> IntConstant -9 loop:none
+ /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none
+ /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<I10>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Add2:i\d+>> Add [<<Get2>>,<<M9>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Add3:i\d+>> Add [<<Add1>>,<<Add2>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add3>>,<<I1>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none
+ //
+ /// CHECK-START-ARM64: void Main.rounding_halving_add_signed_alt2(short[], short[], short[]) loop_optimization (after)
+ /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none
+ /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:true loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none
+ private static void rounding_halving_add_signed_alt2(short[] b1, short[] b2, short[] bo) {
+ int min_length = Math.min(bo.length, Math.min(b1.length, b2.length));
+ for (int i = 0; i < min_length; i++) {
+ // Computations that cancel to adding 1 also do not confuse recognition.
+ bo[i] = (short) (((b1[i] + 10) + (b2[i] - 9)) >> 1);
+ }
+ }
+
/// CHECK-START: void Main.rounding_halving_add_unsigned(short[], short[], short[]) loop_optimization (before)
/// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none
/// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none
@@ -126,6 +207,34 @@ public class Main {
}
}
+ /// CHECK-START: void Main.rounding_halving_add_unsigned_alt(short[], short[], short[]) loop_optimization (before)
+ /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none
+ /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none
+ /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none
+ /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<And1:i\d+>> And [<<Get1>>,<<UMAX>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<And2:i\d+>> And [<<Get2>>,<<UMAX>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Add1:i\d+>> Add [<<And2>>,<<I1>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Add2:i\d+>> Add [<<And1>>,<<Add1>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add2>>,<<I1>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none
+ //
+ /// CHECK-START-ARM64: void Main.rounding_halving_add_unsigned_alt(short[], short[], short[]) loop_optimization (after)
+ /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none
+ /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:true rounded:true loop:<<Loop>> outer_loop:none
+ /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none
+ private static void rounding_halving_add_unsigned_alt(short[] b1, short[] b2, short[] bo) {
+ int min_length = Math.min(bo.length, Math.min(b1.length, b2.length));
+ for (int i = 0; i < min_length; i++) {
+ // Slightly different order in idiom does not confuse recognition.
+ bo[i] = (short) ((b1[i] & 0xffff) + ((b2[i] & 0xffff) + 1) >> 1);
+ }
+ }
+
/// CHECK-START: void Main.halving_add_signed_constant(short[], short[]) loop_optimization (before)
/// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none
/// CHECK-DAG: <<SMAX:i\d+>> IntConstant 32767 loop:none
@@ -200,6 +309,11 @@ public class Main {
short e = (short) ((sB1[i] + sB2[i]) >> 1);
expectEquals(e, sBo[i]);
}
+ halving_add_signed_alt(sB1, sB2, sBo);
+ for (int i = 0; i < M; i++) {
+ short e = (short) ((sB1[i] + sB2[i]) >> 1);
+ expectEquals(e, sBo[i]);
+ }
halving_add_unsigned(sB1, sB2, sBo);
for (int i = 0; i < M; i++) {
short e = (short) (((sB1[i] & 0xffff) + (sB2[i] & 0xffff)) >> 1);
@@ -210,11 +324,26 @@ public class Main {
short e = (short) ((sB1[i] + sB2[i] + 1) >> 1);
expectEquals(e, sBo[i]);
}
+ rounding_halving_add_signed_alt(sB1, sB2, sBo);
+ for (int i = 0; i < M; i++) {
+ short e = (short) ((sB1[i] + sB2[i] + 1) >> 1);
+ expectEquals(e, sBo[i]);
+ }
+ rounding_halving_add_signed_alt2(sB1, sB2, sBo);
+ for (int i = 0; i < M; i++) {
+ short e = (short) ((sB1[i] + sB2[i] + 1) >> 1);
+ expectEquals(e, sBo[i]);
+ }
rounding_halving_add_unsigned(sB1, sB2, sBo);
for (int i = 0; i < M; i++) {
short e = (short) (((sB1[i] & 0xffff) + (sB2[i] & 0xffff) + 1) >> 1);
expectEquals(e, sBo[i]);
}
+ rounding_halving_add_unsigned_alt(sB1, sB2, sBo);
+ for (int i = 0; i < M; i++) {
+ short e = (short) (((sB1[i] & 0xffff) + (sB2[i] & 0xffff) + 1) >> 1);
+ expectEquals(e, sBo[i]);
+ }
halving_add_signed_constant(sB1, sBo);
for (int i = 0; i < M; i++) {
short e = (short) ((sB1[i] + 0x7fff) >> 1);
generated by cgit v1.2.3-59-g8ed1b (git 2.39.0) at 2026-01-04 04:39:54 +0000