summaryrefslogtreecommitdiff
path: root/compiler/optimizing
diff options
context:
space:
mode:
Diffstat (limited to 'compiler/optimizing')
-rw-r--r--compiler/optimizing/code_generator_x86.cc158
-rw-r--r--compiler/optimizing/code_generator_x86.h4
2 files changed, 74 insertions, 88 deletions
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index ef896aa949..92b62e2c84 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -1525,8 +1525,10 @@ void LocationsBuilderX86::VisitTypeConversion(HTypeConversion* conversion) {
case Primitive::kPrimLong:
// Processing a Dex `long-to-float' instruction.
- locations->SetInAt(0, Location::Any());
- locations->SetOut(Location::Any());
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetOut(Location::RequiresFpuRegister());
+ locations->AddTemp(Location::RequiresFpuRegister());
+ locations->AddTemp(Location::RequiresFpuRegister());
break;
case Primitive::kPrimDouble:
@@ -1556,8 +1558,10 @@ void LocationsBuilderX86::VisitTypeConversion(HTypeConversion* conversion) {
case Primitive::kPrimLong:
// Processing a Dex `long-to-double' instruction.
- locations->SetInAt(0, Location::Any());
- locations->SetOut(Location::Any());
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetOut(Location::RequiresFpuRegister());
+ locations->AddTemp(Location::RequiresFpuRegister());
+ locations->AddTemp(Location::RequiresFpuRegister());
break;
case Primitive::kPrimFloat:
@@ -1778,31 +1782,37 @@ void InstructionCodeGeneratorX86::VisitTypeConversion(HTypeConversion* conversio
case Primitive::kPrimLong: {
// Processing a Dex `long-to-float' instruction.
- size_t adjustment = 0;
-
- // Create stack space for the call to
- // InstructionCodeGeneratorX86::PushOntoFPStack and/or X86Assembler::fstps below.
- // TODO: enhance register allocator to ask for stack temporaries.
- if (!in.IsDoubleStackSlot() || !out.IsStackSlot()) {
- adjustment = Primitive::ComponentSize(Primitive::kPrimLong);
- __ subl(ESP, Immediate(adjustment));
- }
-
- // Load the value to the FP stack, using temporaries if needed.
- PushOntoFPStack(in, 0, adjustment, false, true);
-
- if (out.IsStackSlot()) {
- __ fstps(Address(ESP, out.GetStackIndex() + adjustment));
- } else {
- __ fstps(Address(ESP, 0));
- Location stack_temp = Location::StackSlot(0);
- codegen_->Move32(out, stack_temp);
- }
-
- // Remove the temporary stack space we allocated.
- if (adjustment != 0) {
- __ addl(ESP, Immediate(adjustment));
- }
+ Register low = in.AsRegisterPairLow<Register>();
+ Register high = in.AsRegisterPairHigh<Register>();
+ XmmRegister result = out.AsFpuRegister<XmmRegister>();
+ XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
+ XmmRegister constant = locations->GetTemp(1).AsFpuRegister<XmmRegister>();
+
+ // Operations use doubles for precision reasons (each 32-bit
+ // half of a long fits in the 53-bit mantissa of a double,
+ // but not in the 24-bit mantissa of a float). This is
+ // especially important for the low bits. The result is
+ // eventually converted to float.
+
+ // low = low - 2^31 (to prevent bit 31 of `low` to be
+ // interpreted as a sign bit)
+ __ subl(low, Immediate(0x80000000));
+ // temp = int-to-double(high)
+ __ cvtsi2sd(temp, high);
+ // temp = temp * 2^32
+ __ LoadLongConstant(constant, k2Pow32EncodingForDouble);
+ __ mulsd(temp, constant);
+ // result = int-to-double(low)
+ __ cvtsi2sd(result, low);
+ // result = result + 2^31 (restore the original value of `low`)
+ __ LoadLongConstant(constant, k2Pow31EncodingForDouble);
+ __ addsd(result, constant);
+ // result = result + temp
+ __ addsd(result, temp);
+ // result = double-to-float(result)
+ __ cvtsd2ss(result, result);
+ // Restore low.
+ __ addl(low, Immediate(0x80000000));
break;
}
@@ -1831,31 +1841,29 @@ void InstructionCodeGeneratorX86::VisitTypeConversion(HTypeConversion* conversio
case Primitive::kPrimLong: {
// Processing a Dex `long-to-double' instruction.
- size_t adjustment = 0;
-
- // Create stack space for the call to
- // InstructionCodeGeneratorX86::PushOntoFPStack and/or X86Assembler::fstpl below.
- // TODO: enhance register allocator to ask for stack temporaries.
- if (!in.IsDoubleStackSlot() || !out.IsDoubleStackSlot()) {
- adjustment = Primitive::ComponentSize(Primitive::kPrimLong);
- __ subl(ESP, Immediate(adjustment));
- }
-
- // Load the value to the FP stack, using temporaries if needed.
- PushOntoFPStack(in, 0, adjustment, false, true);
-
- if (out.IsDoubleStackSlot()) {
- __ fstpl(Address(ESP, out.GetStackIndex() + adjustment));
- } else {
- __ fstpl(Address(ESP, 0));
- Location stack_temp = Location::DoubleStackSlot(0);
- codegen_->Move64(out, stack_temp);
- }
-
- // Remove the temporary stack space we allocated.
- if (adjustment != 0) {
- __ addl(ESP, Immediate(adjustment));
- }
+ Register low = in.AsRegisterPairLow<Register>();
+ Register high = in.AsRegisterPairHigh<Register>();
+ XmmRegister result = out.AsFpuRegister<XmmRegister>();
+ XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
+ XmmRegister constant = locations->GetTemp(1).AsFpuRegister<XmmRegister>();
+
+ // low = low - 2^31 (to prevent bit 31 of `low` to be
+ // interpreted as a sign bit)
+ __ subl(low, Immediate(0x80000000));
+ // temp = int-to-double(high)
+ __ cvtsi2sd(temp, high);
+ // temp = temp * 2^32
+ __ LoadLongConstant(constant, k2Pow32EncodingForDouble);
+ __ mulsd(temp, constant);
+ // result = int-to-double(low)
+ __ cvtsi2sd(result, low);
+ // result = result + 2^31 (restore the original value of `low`)
+ __ LoadLongConstant(constant, k2Pow31EncodingForDouble);
+ __ addsd(result, constant);
+ // result = result + temp
+ __ addsd(result, temp);
+ // Restore low.
+ __ addl(low, Immediate(0x80000000));
break;
}
@@ -2195,43 +2203,24 @@ void InstructionCodeGeneratorX86::VisitMul(HMul* mul) {
}
}
-void InstructionCodeGeneratorX86::PushOntoFPStack(Location source,
- uint32_t temp_offset,
- uint32_t stack_adjustment,
- bool is_fp,
- bool is_wide) {
+void InstructionCodeGeneratorX86::PushOntoFPStack(Location source, uint32_t temp_offset,
+ uint32_t stack_adjustment, bool is_float) {
if (source.IsStackSlot()) {
- DCHECK(!is_wide);
- if (is_fp) {
- __ flds(Address(ESP, source.GetStackIndex() + stack_adjustment));
- } else {
- __ filds(Address(ESP, source.GetStackIndex() + stack_adjustment));
- }
+ DCHECK(is_float);
+ __ flds(Address(ESP, source.GetStackIndex() + stack_adjustment));
} else if (source.IsDoubleStackSlot()) {
- DCHECK(is_wide);
- if (is_fp) {
- __ fldl(Address(ESP, source.GetStackIndex() + stack_adjustment));
- } else {
- __ fildl(Address(ESP, source.GetStackIndex() + stack_adjustment));
- }
+ DCHECK(!is_float);
+ __ fldl(Address(ESP, source.GetStackIndex() + stack_adjustment));
} else {
// Write the value to the temporary location on the stack and load to FP stack.
- if (!is_wide) {
+ if (is_float) {
Location stack_temp = Location::StackSlot(temp_offset);
codegen_->Move32(stack_temp, source);
- if (is_fp) {
- __ flds(Address(ESP, temp_offset));
- } else {
- __ filds(Address(ESP, temp_offset));
- }
+ __ flds(Address(ESP, temp_offset));
} else {
Location stack_temp = Location::DoubleStackSlot(temp_offset);
codegen_->Move64(stack_temp, source);
- if (is_fp) {
- __ fldl(Address(ESP, temp_offset));
- } else {
- __ fildl(Address(ESP, temp_offset));
- }
+ __ fldl(Address(ESP, temp_offset));
}
}
}
@@ -2250,9 +2239,8 @@ void InstructionCodeGeneratorX86::GenerateRemFP(HRem *rem) {
__ subl(ESP, Immediate(2 * elem_size));
// Load the values to the FP stack in reverse order, using temporaries if needed.
- const bool is_wide = !is_float;
- PushOntoFPStack(second, elem_size, 2 * elem_size, /* is_fp */ true, is_wide);
- PushOntoFPStack(first, 0, 2 * elem_size, /* is_fp */ true, is_wide);
+ PushOntoFPStack(second, elem_size, 2 * elem_size, is_float);
+ PushOntoFPStack(first, 0, 2 * elem_size, is_float);
// Loop doing FPREM until we stabilize.
Label retry;
diff --git a/compiler/optimizing/code_generator_x86.h b/compiler/optimizing/code_generator_x86.h
index 1cb9e326ce..0cc3c6533a 100644
--- a/compiler/optimizing/code_generator_x86.h
+++ b/compiler/optimizing/code_generator_x86.h
@@ -171,10 +171,8 @@ class InstructionCodeGeneratorX86 : public HGraphVisitor {
void GenerateMemoryBarrier(MemBarrierKind kind);
void HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info);
void HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info);
- // Push value to FPU stack. `is_fp` specifies whether the value is floating point or not.
- // `is_wide` specifies whether it is long/double or not.
void PushOntoFPStack(Location source, uint32_t temp_offset,
- uint32_t stack_adjustment, bool is_fp, bool is_wide);
+ uint32_t stack_adjustment, bool is_float);
void GenerateImplicitNullCheck(HNullCheck* instruction);
void GenerateExplicitNullCheck(HNullCheck* instruction);
generated by cgit v1.2.3-59-g8ed1b (git 2.39.0) at 2026-01-04 14:55:00 +0000