Revert "Revert "ART: Split out more cases of Load/StoreRef, volatile as parameter""
This reverts commit de68676b24f61a55adc0b22fe828f036a5925c41.
Fixes an API comment, and differentiates between inserting and appending.
Change-Id: I0e9a21bb1d25766e3cbd802d8b48633ae251a6bf
diff --git a/compiler/dex/quick/arm/call_arm.cc b/compiler/dex/quick/arm/call_arm.cc
index 590c767..04d6898 100644
--- a/compiler/dex/quick/arm/call_arm.cc
+++ b/compiler/dex/quick/arm/call_arm.cc
@@ -316,9 +316,9 @@
int ex_offset = Thread::ExceptionOffset<4>().Int32Value();
RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true);
RegStorage reset_reg = AllocTempRef();
- LoadRefDisp(rs_rARM_SELF, ex_offset, rl_result.reg);
+ LoadRefDisp(rs_rARM_SELF, ex_offset, rl_result.reg, kNotVolatile);
LoadConstant(reset_reg, 0);
- StoreRefDisp(rs_rARM_SELF, ex_offset, reset_reg);
+ StoreRefDisp(rs_rARM_SELF, ex_offset, reset_reg, kNotVolatile);
FreeTemp(reset_reg);
StoreValue(rl_dest, rl_result);
}
diff --git a/compiler/dex/quick/arm/codegen_arm.h b/compiler/dex/quick/arm/codegen_arm.h
index 4499862..70dce7f 100644
--- a/compiler/dex/quick/arm/codegen_arm.h
+++ b/compiler/dex/quick/arm/codegen_arm.h
@@ -33,20 +33,16 @@
LIR* CheckSuspendUsingLoad() OVERRIDE;
RegStorage LoadHelper(ThreadOffset<4> offset) OVERRIDE;
RegStorage LoadHelper(ThreadOffset<8> offset) OVERRIDE;
- LIR* LoadBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_dest,
- OpSize size) OVERRIDE;
LIR* LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
- OpSize size) OVERRIDE;
+ OpSize size, VolatileKind is_volatile) OVERRIDE;
LIR* LoadBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_dest, int scale,
OpSize size) OVERRIDE;
LIR* LoadBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale, int displacement,
RegStorage r_dest, OpSize size) OVERRIDE;
LIR* LoadConstantNoClobber(RegStorage r_dest, int value);
LIR* LoadConstantWide(RegStorage r_dest, int64_t value);
- LIR* StoreBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_src,
- OpSize size) OVERRIDE;
LIR* StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src,
- OpSize size) OVERRIDE;
+ OpSize size, VolatileKind is_volatile) OVERRIDE;
LIR* StoreBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_src, int scale,
OpSize size) OVERRIDE;
LIR* StoreBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale, int displacement,
diff --git a/compiler/dex/quick/arm/int_arm.cc b/compiler/dex/quick/arm/int_arm.cc
index 916c528..e34d944 100644
--- a/compiler/dex/quick/arm/int_arm.cc
+++ b/compiler/dex/quick/arm/int_arm.cc
@@ -723,7 +723,7 @@
} else {
DCHECK(size == kSignedByte || size == kSignedHalf || size == k32);
// Unaligned load with LDR and LDRSH is allowed on ARMv7 with SCTLR.A set to 0.
- LoadBaseDisp(rl_address.reg, 0, rl_result.reg, size);
+ LoadBaseDisp(rl_address.reg, 0, rl_result.reg, size, kNotVolatile);
StoreValue(rl_dest, rl_result);
}
return true;
@@ -737,13 +737,13 @@
if (size == k64) {
// Fake unaligned STRD by two unaligned STR instructions on ARMv7 with SCTLR.A set to 0.
RegLocation rl_value = LoadValueWide(rl_src_value, kCoreReg);
- StoreBaseDisp(rl_address.reg, 0, rl_value.reg.GetLow(), k32);
- StoreBaseDisp(rl_address.reg, 4, rl_value.reg.GetHigh(), k32);
+ StoreBaseDisp(rl_address.reg, 0, rl_value.reg.GetLow(), k32, kNotVolatile);
+ StoreBaseDisp(rl_address.reg, 4, rl_value.reg.GetHigh(), k32, kNotVolatile);
} else {
DCHECK(size == kSignedByte || size == kSignedHalf || size == k32);
// Unaligned store with STR and STRSH is allowed on ARMv7 with SCTLR.A set to 0.
RegLocation rl_value = LoadValue(rl_src_value, kCoreReg);
- StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size);
+ StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size, kNotVolatile);
}
return true;
}
@@ -1230,7 +1230,7 @@
}
FreeTemp(reg_len);
}
- LoadBaseDisp(reg_ptr, data_offset, rl_result.reg, size);
+ LoadBaseDisp(reg_ptr, data_offset, rl_result.reg, size, kNotVolatile);
MarkPossibleNullPointerException(opt_flags);
if (!constant_index) {
FreeTemp(reg_ptr);
@@ -1330,7 +1330,7 @@
FreeTemp(reg_len);
}
- StoreBaseDisp(reg_ptr, data_offset, rl_src.reg, size);
+ StoreBaseDisp(reg_ptr, data_offset, rl_src.reg, size, kNotVolatile);
MarkPossibleNullPointerException(opt_flags);
} else {
/* reg_ptr -> array data */
diff --git a/compiler/dex/quick/arm/utility_arm.cc b/compiler/dex/quick/arm/utility_arm.cc
index b236f99..bc8f95b 100644
--- a/compiler/dex/quick/arm/utility_arm.cc
+++ b/compiler/dex/quick/arm/utility_arm.cc
@@ -961,31 +961,37 @@
return load;
}
-LIR* ArmMir2Lir::LoadBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_dest,
- OpSize size) {
- // Only 64-bit load needs special handling.
- if (UNLIKELY(size == k64 || size == kDouble)) {
- DCHECK(!r_dest.IsFloat()); // See RegClassForFieldLoadSave().
- // If the cpu supports LPAE, aligned LDRD is atomic - fall through to LoadBaseDisp().
- if (!cu_->compiler_driver->GetInstructionSetFeatures().HasLpae()) {
- // Use LDREXD for the atomic load. (Expect displacement > 0, don't optimize for == 0.)
- RegStorage r_ptr = AllocTemp();
- OpRegRegImm(kOpAdd, r_ptr, r_base, displacement);
- LIR* lir = NewLIR3(kThumb2Ldrexd, r_dest.GetLowReg(), r_dest.GetHighReg(), r_ptr.GetReg());
- FreeTemp(r_ptr);
- return lir;
- }
- }
- return LoadBaseDisp(r_base, displacement, r_dest, size);
-}
-
LIR* ArmMir2Lir::LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest,
- OpSize size) {
+ OpSize size, VolatileKind is_volatile) {
// TODO: base this on target.
if (size == kWord) {
size = k32;
}
- return LoadBaseDispBody(r_base, displacement, r_dest, size);
+ LIR* load;
+ if (UNLIKELY(is_volatile == kVolatile &&
+ (size == k64 || size == kDouble) &&
+ !cu_->compiler_driver->GetInstructionSetFeatures().HasLpae())) {
+ // Only 64-bit load needs special handling.
+ // If the cpu supports LPAE, aligned LDRD is atomic - fall through to LoadBaseDisp().
+ DCHECK(!r_dest.IsFloat()); // See RegClassForFieldLoadSave().
+ // Use LDREXD for the atomic load. (Expect displacement > 0, don't optimize for == 0.)
+ RegStorage r_ptr = AllocTemp();
+ OpRegRegImm(kOpAdd, r_ptr, r_base, displacement);
+ LIR* lir = NewLIR3(kThumb2Ldrexd, r_dest.GetLowReg(), r_dest.GetHighReg(), r_ptr.GetReg());
+ FreeTemp(r_ptr);
+ return lir;
+ } else {
+ load = LoadBaseDispBody(r_base, displacement, r_dest, size);
+ }
+
+ if (UNLIKELY(is_volatile == kVolatile)) {
+ // Without context sensitive analysis, we must issue the most conservative barriers.
+ // In this case, either a load or store may follow so we issue both barriers.
+ GenMemBarrier(kLoadLoad);
+ GenMemBarrier(kLoadStore);
+ }
+
+ return load;
}
@@ -1081,49 +1087,58 @@
return store;
}
-LIR* ArmMir2Lir::StoreBaseDispVolatile(RegStorage r_base, int displacement, RegStorage r_src,
- OpSize size) {
- // Only 64-bit store needs special handling.
- if (UNLIKELY(size == k64 || size == kDouble)) {
- DCHECK(!r_src.IsFloat()); // See RegClassForFieldLoadSave().
- // If the cpu supports LPAE, aligned STRD is atomic - fall through to StoreBaseDisp().
- if (!cu_->compiler_driver->GetInstructionSetFeatures().HasLpae()) {
- // Use STREXD for the atomic store. (Expect displacement > 0, don't optimize for == 0.)
- RegStorage r_ptr = AllocTemp();
- OpRegRegImm(kOpAdd, r_ptr, r_base, displacement);
- LIR* fail_target = NewLIR0(kPseudoTargetLabel);
- // We have only 5 temporary registers available and if r_base, r_src and r_ptr already
- // take 4, we can't directly allocate 2 more for LDREXD temps. In that case clobber r_ptr
- // in LDREXD and recalculate it from r_base.
- RegStorage r_temp = AllocTemp();
- RegStorage r_temp_high = AllocFreeTemp(); // We may not have another temp.
- if (r_temp_high.Valid()) {
- NewLIR3(kThumb2Ldrexd, r_temp.GetReg(), r_temp_high.GetReg(), r_ptr.GetReg());
- FreeTemp(r_temp_high);
- FreeTemp(r_temp);
- } else {
- // If we don't have another temp, clobber r_ptr in LDREXD and reload it.
- NewLIR3(kThumb2Ldrexd, r_temp.GetReg(), r_ptr.GetReg(), r_ptr.GetReg());
- FreeTemp(r_temp); // May need the temp for kOpAdd.
- OpRegRegImm(kOpAdd, r_ptr, r_base, displacement);
- }
- LIR* lir = NewLIR4(kThumb2Strexd, r_temp.GetReg(), r_src.GetLowReg(), r_src.GetHighReg(),
- r_ptr.GetReg());
- OpCmpImmBranch(kCondNe, r_temp, 0, fail_target);
- FreeTemp(r_ptr);
- return lir;
- }
- }
- return StoreBaseDisp(r_base, displacement, r_src, size);
-}
-
LIR* ArmMir2Lir::StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src,
- OpSize size) {
- // TODO: base this on target.
- if (size == kWord) {
- size = k32;
+ OpSize size, VolatileKind is_volatile) {
+ if (UNLIKELY(is_volatile == kVolatile)) {
+ // There might have been a store before this volatile one so insert StoreStore barrier.
+ GenMemBarrier(kStoreStore);
}
- return StoreBaseDispBody(r_base, displacement, r_src, size);
+
+ LIR* store;
+ if (UNLIKELY(is_volatile == kVolatile &&
+ (size == k64 || size == kDouble) &&
+ !cu_->compiler_driver->GetInstructionSetFeatures().HasLpae())) {
+ // Only 64-bit store needs special handling.
+ // If the cpu supports LPAE, aligned STRD is atomic - fall through to StoreBaseDisp().
+ // Use STREXD for the atomic store. (Expect displacement > 0, don't optimize for == 0.)
+ DCHECK(!r_src.IsFloat()); // See RegClassForFieldLoadSave().
+ RegStorage r_ptr = AllocTemp();
+ OpRegRegImm(kOpAdd, r_ptr, r_base, displacement);
+ LIR* fail_target = NewLIR0(kPseudoTargetLabel);
+ // We have only 5 temporary registers available and if r_base, r_src and r_ptr already
+ // take 4, we can't directly allocate 2 more for LDREXD temps. In that case clobber r_ptr
+ // in LDREXD and recalculate it from r_base.
+ RegStorage r_temp = AllocTemp();
+ RegStorage r_temp_high = AllocFreeTemp(); // We may not have another temp.
+ if (r_temp_high.Valid()) {
+ NewLIR3(kThumb2Ldrexd, r_temp.GetReg(), r_temp_high.GetReg(), r_ptr.GetReg());
+ FreeTemp(r_temp_high);
+ FreeTemp(r_temp);
+ } else {
+ // If we don't have another temp, clobber r_ptr in LDREXD and reload it.
+ NewLIR3(kThumb2Ldrexd, r_temp.GetReg(), r_ptr.GetReg(), r_ptr.GetReg());
+ FreeTemp(r_temp); // May need the temp for kOpAdd.
+ OpRegRegImm(kOpAdd, r_ptr, r_base, displacement);
+ }
+ store = NewLIR4(kThumb2Strexd, r_temp.GetReg(), r_src.GetLowReg(), r_src.GetHighReg(),
+ r_ptr.GetReg());
+ OpCmpImmBranch(kCondNe, r_temp, 0, fail_target);
+ FreeTemp(r_ptr);
+ } else {
+ // TODO: base this on target.
+ if (size == kWord) {
+ size = k32;
+ }
+
+ store = StoreBaseDispBody(r_base, displacement, r_src, size);
+ }
+
+ if (UNLIKELY(is_volatile == kVolatile)) {
+ // A load might follow the volatile store so insert a StoreLoad barrier.
+ GenMemBarrier(kStoreLoad);
+ }
+
+ return store;
}
LIR* ArmMir2Lir::OpFpRegCopy(RegStorage r_dest, RegStorage r_src) {