Optimize double/float immediate loading on arm.
Also reserve a D register for temp.
Change-Id: I6584d9005b0f5685c3afcd8e9153b4c87b56aa8e
diff --git a/compiler/optimizing/code_generator_arm.cc b/compiler/optimizing/code_generator_arm.cc
index e5de2ab..2a79f82 100644
--- a/compiler/optimizing/code_generator_arm.cc
+++ b/compiler/optimizing/code_generator_arm.cc
@@ -55,6 +55,10 @@
static constexpr SRegister kFpuCalleeSaves[] =
{ S16, S17, S18, S19, S20, S21, S22, S23, S24, S25, S26, S27, S28, S29, S30, S31 };
+// D31 cannot be split into two S registers, and the register allocator only works on
+// S registers. Therefore there is no need to block it.
+static constexpr DRegister DTMP = D31;
+
class InvokeRuntimeCallingConvention : public CallingConvention<Register, SRegister> {
public:
InvokeRuntimeCallingConvention()
@@ -3361,10 +3365,8 @@
}
} else if (source.IsDoubleStackSlot()) {
if (destination.IsDoubleStackSlot()) {
- __ LoadFromOffset(kLoadWord, IP, SP, source.GetStackIndex());
- __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex());
- __ LoadFromOffset(kLoadWord, IP, SP, source.GetHighStackIndex(kArmWordSize));
- __ StoreToOffset(kStoreWord, IP, SP, destination.GetHighStackIndex(kArmWordSize));
+ __ LoadDFromOffset(DTMP, SP, source.GetStackIndex());
+ __ StoreDToOffset(DTMP, SP, destination.GetStackIndex());
} else if (destination.IsRegisterPair()) {
DCHECK(ExpectedPairLayout(destination));
__ LoadFromOffset(
@@ -3484,16 +3486,13 @@
__ vmovs(source.AsFpuRegister<SRegister>(), destination.AsFpuRegister<SRegister>());
__ vmovsr(destination.AsFpuRegister<SRegister>(), IP);
} else if (source.IsRegisterPair() && destination.IsRegisterPair()) {
- __ Mov(IP, source.AsRegisterPairLow<Register>());
+ __ vmovdrr(DTMP, source.AsRegisterPairLow<Register>(), source.AsRegisterPairHigh<Register>());
__ Mov(source.AsRegisterPairLow<Register>(), destination.AsRegisterPairLow<Register>());
- __ Mov(destination.AsRegisterPairLow<Register>(), IP);
- __ Mov(IP, source.AsRegisterPairHigh<Register>());
__ Mov(source.AsRegisterPairHigh<Register>(), destination.AsRegisterPairHigh<Register>());
- __ Mov(destination.AsRegisterPairHigh<Register>(), IP);
+ __ vmovrrd(destination.AsRegisterPairLow<Register>(),
+ destination.AsRegisterPairHigh<Register>(),
+ DTMP);
} else if (source.IsRegisterPair() || destination.IsRegisterPair()) {
- // TODO: Find a D register available in the parallel moves,
- // or reserve globally a D register.
- DRegister tmp = D0;
Register low_reg = source.IsRegisterPair()
? source.AsRegisterPairLow<Register>()
: destination.AsRegisterPairLow<Register>();
@@ -3501,27 +3500,15 @@
? destination.GetStackIndex()
: source.GetStackIndex();
DCHECK(ExpectedPairLayout(source.IsRegisterPair() ? source : destination));
- // Make room for the pushed DRegister.
- mem += 8;
- __ vpushd(tmp, 1);
- __ vmovdrr(tmp, low_reg, static_cast<Register>(low_reg + 1));
+ __ vmovdrr(DTMP, low_reg, static_cast<Register>(low_reg + 1));
__ LoadFromOffset(kLoadWordPair, low_reg, SP, mem);
- __ StoreDToOffset(tmp, SP, mem);
- __ vpopd(tmp, 1);
+ __ StoreDToOffset(DTMP, SP, mem);
} else if (source.IsFpuRegisterPair() && destination.IsFpuRegisterPair()) {
- // TODO: Find a D register available in the parallel moves,
- // or reserve globally a D register.
- DRegister tmp = D0;
DRegister first = FromLowSToD(source.AsFpuRegisterPairLow<SRegister>());
DRegister second = FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>());
- while (tmp == first || tmp == second) {
- tmp = static_cast<DRegister>(tmp + 1);
- }
- __ vpushd(tmp, 1);
- __ vmovd(tmp, first);
+ __ vmovd(DTMP, first);
__ vmovd(first, second);
- __ vmovd(second, tmp);
- __ vpopd(tmp, 1);
+ __ vmovd(second, DTMP);
} else if (source.IsFpuRegisterPair() || destination.IsFpuRegisterPair()) {
DRegister reg = source.IsFpuRegisterPair()
? FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())
@@ -3529,15 +3516,9 @@
int mem = source.IsFpuRegisterPair()
? destination.GetStackIndex()
: source.GetStackIndex();
- // TODO: Find or reserve a D register.
- DRegister tmp = reg == D0 ? D1 : D0;
- // Make room for the pushed DRegister.
- mem += 8;
- __ vpushd(tmp, 1);
- __ vmovd(tmp, reg);
+ __ vmovd(DTMP, reg);
__ LoadDFromOffset(reg, SP, mem);
- __ StoreDToOffset(tmp, SP, mem);
- __ vpopd(tmp, 1);
+ __ StoreDToOffset(DTMP, SP, mem);
} else if (source.IsFpuRegister() || destination.IsFpuRegister()) {
SRegister reg = source.IsFpuRegister() ? source.AsFpuRegister<SRegister>()
: destination.AsFpuRegister<SRegister>();
diff --git a/compiler/utils/arm/assembler_arm.h b/compiler/utils/arm/assembler_arm.h
index 0d84ba7..8730f52 100644
--- a/compiler/utils/arm/assembler_arm.h
+++ b/compiler/utils/arm/assembler_arm.h
@@ -536,18 +536,44 @@
virtual void LoadImmediate(Register rd, int32_t value, Condition cond = AL) = 0;
void LoadSImmediate(SRegister sd, float value, Condition cond = AL) {
if (!vmovs(sd, value, cond)) {
- LoadImmediate(IP, bit_cast<int32_t, float>(value), cond);
- vmovsr(sd, IP, cond);
+ int32_t int_value = bit_cast<int32_t, float>(value);
+ if (int_value == bit_cast<int32_t, float>(0.0f)) {
+ // 0.0 is quite common, so we special case it by loading
+ // 2.0 in `sd` and then substracting it.
+ bool success = vmovs(sd, 2.0, cond);
+ CHECK(success);
+ vsubs(sd, sd, sd, cond);
+ } else {
+ LoadImmediate(IP, int_value, cond);
+ vmovsr(sd, IP, cond);
+ }
}
}
void LoadDImmediate(DRegister sd, double value, Condition cond = AL) {
if (!vmovd(sd, value, cond)) {
uint64_t int_value = bit_cast<uint64_t, double>(value);
- LoadSImmediate(
- static_cast<SRegister>(sd << 1), bit_cast<float, uint32_t>(Low32Bits(int_value)));
- LoadSImmediate(
- static_cast<SRegister>((sd << 1) + 1), bit_cast<float, uint32_t>(High32Bits(int_value)));
+ if (int_value == bit_cast<uint64_t, double>(0.0)) {
+ // 0.0 is quite common, so we special case it by loading
+ // 2.0 in `sd` and then substracting it.
+ bool success = vmovd(sd, 2.0, cond);
+ CHECK(success);
+ vsubd(sd, sd, sd, cond);
+ } else {
+ if (sd < 16) {
+ SRegister low = static_cast<SRegister>(sd << 1);
+ SRegister high = static_cast<SRegister>(low + 1);
+ LoadSImmediate(low, bit_cast<float, uint32_t>(Low32Bits(int_value)), cond);
+ if (High32Bits(int_value) == Low32Bits(int_value)) {
+ vmovs(high, low);
+ } else {
+ LoadSImmediate(high, bit_cast<float, uint32_t>(High32Bits(int_value)), cond);
+ }
+ } else {
+ LOG(FATAL) << "Unimplemented loading of double into a D register "
+ << "that cannot be split into two S registers";
+ }
+ }
}
}