x86/x86-64: Shorter fast-path for read barrier field load.
Test: Tested with ART_USE_READ_BARRIER=true on host.
Bug: 29966877
Bug: 12687968
Change-Id: I73359495910dacb2cc28f1a21ef9e610bab5a476
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index ade2117..a85cd54 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -7099,12 +7099,6 @@
// /* LockWord */ lock_word = LockWord(monitor)
static_assert(sizeof(LockWord) == sizeof(int32_t),
"art::LockWord and int32_t have different sizes.");
- // /* uint32_t */ rb_state = lock_word.ReadBarrierState()
- __ shrl(temp_reg, Immediate(LockWord::kReadBarrierStateShift));
- __ andl(temp_reg, Immediate(LockWord::kReadBarrierStateMask));
- static_assert(
- LockWord::kReadBarrierStateMask == ReadBarrier::rb_ptr_mask_,
- "art::LockWord::kReadBarrierStateMask is not equal to art::ReadBarrier::rb_ptr_mask_.");
// Load fence to prevent load-load reordering.
// Note that this is a no-op, thanks to the x86 memory model.
@@ -7124,8 +7118,13 @@
// if (rb_state == ReadBarrier::gray_ptr_)
// ref = ReadBarrier::Mark(ref);
- __ cmpl(temp_reg, Immediate(ReadBarrier::gray_ptr_));
- __ j(kEqual, slow_path->GetEntryLabel());
+ // Given the numeric representation, it's enough to check the low bit of the
+ // rb_state. We do that by shifting the bit out of the lock word with SHR.
+ static_assert(ReadBarrier::white_ptr_ == 0, "Expecting white to have value 0");
+ static_assert(ReadBarrier::gray_ptr_ == 1, "Expecting gray to have value 1");
+ static_assert(ReadBarrier::black_ptr_ == 2, "Expecting black to have value 2");
+ __ shrl(temp_reg, Immediate(LockWord::kReadBarrierStateShift + 1));
+ __ j(kCarrySet, slow_path->GetEntryLabel());
__ Bind(slow_path->GetExitLabel());
}
diff --git a/compiler/optimizing/code_generator_x86_64.cc b/compiler/optimizing/code_generator_x86_64.cc
index eadb431..e001363 100644
--- a/compiler/optimizing/code_generator_x86_64.cc
+++ b/compiler/optimizing/code_generator_x86_64.cc
@@ -6551,12 +6551,6 @@
// /* LockWord */ lock_word = LockWord(monitor)
static_assert(sizeof(LockWord) == sizeof(int32_t),
"art::LockWord and int32_t have different sizes.");
- // /* uint32_t */ rb_state = lock_word.ReadBarrierState()
- __ shrl(temp_reg, Immediate(LockWord::kReadBarrierStateShift));
- __ andl(temp_reg, Immediate(LockWord::kReadBarrierStateMask));
- static_assert(
- LockWord::kReadBarrierStateMask == ReadBarrier::rb_ptr_mask_,
- "art::LockWord::kReadBarrierStateMask is not equal to art::ReadBarrier::rb_ptr_mask_.");
// Load fence to prevent load-load reordering.
// Note that this is a no-op, thanks to the x86-64 memory model.
@@ -6576,8 +6570,13 @@
// if (rb_state == ReadBarrier::gray_ptr_)
// ref = ReadBarrier::Mark(ref);
- __ cmpl(temp_reg, Immediate(ReadBarrier::gray_ptr_));
- __ j(kEqual, slow_path->GetEntryLabel());
+ // Given the numeric representation, it's enough to check the low bit of the
+ // rb_state. We do that by shifting the bit out of the lock word with SHR.
+ static_assert(ReadBarrier::white_ptr_ == 0, "Expecting white to have value 0");
+ static_assert(ReadBarrier::gray_ptr_ == 1, "Expecting gray to have value 1");
+ static_assert(ReadBarrier::black_ptr_ == 2, "Expecting black to have value 2");
+ __ shrl(temp_reg, Immediate(LockWord::kReadBarrierStateShift + 1));
+ __ j(kCarrySet, slow_path->GetEntryLabel());
__ Bind(slow_path->GetExitLabel());
}
diff --git a/compiler/utils/x86/constants_x86.h b/compiler/utils/x86/constants_x86.h
index 2dfb65c..0bc1560 100644
--- a/compiler/utils/x86/constants_x86.h
+++ b/compiler/utils/x86/constants_x86.h
@@ -97,6 +97,8 @@
kNotZero = kNotEqual,
kNegative = kSign,
kPositive = kNotSign,
+ kCarrySet = kBelow,
+ kCarryClear = kAboveEqual,
kUnordered = kParityEven
};
diff --git a/compiler/utils/x86_64/constants_x86_64.h b/compiler/utils/x86_64/constants_x86_64.h
index 37db6b1..cc508a1 100644
--- a/compiler/utils/x86_64/constants_x86_64.h
+++ b/compiler/utils/x86_64/constants_x86_64.h
@@ -106,6 +106,8 @@
kNotZero = kNotEqual,
kNegative = kSign,
kPositive = kNotSign,
+ kCarrySet = kBelow,
+ kCarryClear = kAboveEqual,
kUnordered = kParityEven
};