ARM64: Instruction simplification for array accesses.
HArrayGet and HArraySet with variable indexes generate two
instructions on arm64, like
add temp, obj, #data_offset
ldr out, [temp, index LSL #shift_amount]
When we have multiple accesses to the same array, the initial `add`
instruction is redundant.
This patch introduces the first instruction simplification in the
arm64-specific instruction simplification pass. It splits HArrayGet
and HArraySet using the new arm64-specific IR HIntermediateAddress.
After that we run GVN again to squash the multiple occurrences of
HIntermediateAddress.
Change-Id: I2e3d12fbb07fed07b2cb2f3f47f99f5a032f8312
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h
index 939e62c..490bfc8 100644
--- a/compiler/optimizing/nodes.h
+++ b/compiler/optimizing/nodes.h
@@ -1079,14 +1079,23 @@
#define FOR_EACH_CONCRETE_INSTRUCTION_ARM(M)
+#ifndef ART_ENABLE_CODEGEN_arm64
#define FOR_EACH_CONCRETE_INSTRUCTION_ARM64(M)
+#else
+#define FOR_EACH_CONCRETE_INSTRUCTION_ARM64(M) \
+ M(Arm64IntermediateAddress, Instruction)
+#endif
#define FOR_EACH_CONCRETE_INSTRUCTION_MIPS64(M)
+#ifndef ART_ENABLE_CODEGEN_x86
+#define FOR_EACH_CONCRETE_INSTRUCTION_X86(M)
+#else
#define FOR_EACH_CONCRETE_INSTRUCTION_X86(M) \
M(X86ComputeBaseMethodAddress, Instruction) \
M(X86LoadFromConstantTable, Instruction) \
M(X86PackedSwitch, Instruction)
+#endif
#define FOR_EACH_CONCRETE_INSTRUCTION_X86_64(M)
@@ -1370,6 +1379,10 @@
return SideEffects(flags_ & ~other.flags_);
}
+ void Add(SideEffects other) {
+ flags_ |= other.flags_;
+ }
+
bool Includes(SideEffects other) const {
return (other.flags_ & flags_) == other.flags_;
}
@@ -1943,6 +1956,7 @@
}
SideEffects GetSideEffects() const { return side_effects_; }
+ void AddSideEffects(SideEffects other) { side_effects_.Add(other); }
size_t GetLifetimePosition() const { return lifetime_position_; }
void SetLifetimePosition(size_t position) { lifetime_position_ = position; }
@@ -2012,7 +2026,7 @@
// order of blocks where this instruction's live interval start.
size_t lifetime_position_;
- const SideEffects side_effects_;
+ SideEffects side_effects_;
// TODO: for primitive types this should be marked as invalid.
ReferenceTypeInfo reference_type_info_;
@@ -4424,8 +4438,11 @@
HArrayGet(HInstruction* array,
HInstruction* index,
Primitive::Type type,
- uint32_t dex_pc)
- : HExpression(type, SideEffects::ArrayReadOfType(type), dex_pc) {
+ uint32_t dex_pc,
+ SideEffects additional_side_effects = SideEffects::None())
+ : HExpression(type,
+ SideEffects::ArrayReadOfType(type).Union(additional_side_effects),
+ dex_pc) {
SetRawInputAt(0, array);
SetRawInputAt(1, index);
}
@@ -4460,10 +4477,13 @@
HInstruction* index,
HInstruction* value,
Primitive::Type expected_component_type,
- uint32_t dex_pc)
+ uint32_t dex_pc,
+ SideEffects additional_side_effects = SideEffects::None())
: HTemplateInstruction(
SideEffects::ArrayWriteOfType(expected_component_type).Union(
- SideEffectsForArchRuntimeCalls(value->GetType())), dex_pc),
+ SideEffectsForArchRuntimeCalls(value->GetType())).Union(
+ additional_side_effects),
+ dex_pc),
expected_component_type_(expected_component_type),
needs_type_check_(value->GetType() == Primitive::kPrimNot),
value_can_be_null_(true),
@@ -4518,6 +4538,10 @@
: expected_component_type_;
}
+ Primitive::Type GetRawExpectedComponentType() const {
+ return expected_component_type_;
+ }
+
static SideEffects SideEffectsForArchRuntimeCalls(Primitive::Type value_type) {
return (value_type == Primitive::kPrimNot) ? SideEffects::CanTriggerGC() : SideEffects::None();
}
@@ -4576,6 +4600,7 @@
bool CanThrow() const OVERRIDE { return true; }
+ HInstruction* GetIndex() const { return InputAt(0); }
DECLARE_INSTRUCTION(BoundsCheck);
@@ -5389,6 +5414,9 @@
} // namespace art
+#ifdef ART_ENABLE_CODEGEN_arm64
+#include "nodes_arm64.h"
+#endif
#ifdef ART_ENABLE_CODEGEN_x86
#include "nodes_x86.h"
#endif