Support float & double negation in the optimizing compiler.
- Add support for the neg-float and neg-double Dex
instructions in the optimizing compiler.
- Generate x86, x86-64 and ARM (but not ARM64) code for
float and double HNeg nodes.
- Add related tests to test/415-optimizing-arith-neg.
Change-Id: I29739a86e13dbe6f64e191641d01637c867cba6c
diff --git a/compiler/optimizing/builder.cc b/compiler/optimizing/builder.cc
index e43841a..a4af900 100644
--- a/compiler/optimizing/builder.cc
+++ b/compiler/optimizing/builder.cc
@@ -896,6 +896,16 @@
break;
}
+ case Instruction::NEG_FLOAT: {
+ Unop_12x<HNeg>(instruction, Primitive::kPrimFloat);
+ break;
+ }
+
+ case Instruction::NEG_DOUBLE: {
+ Unop_12x<HNeg>(instruction, Primitive::kPrimDouble);
+ break;
+ }
+
case Instruction::NOT_INT: {
Unop_12x<HNot>(instruction, Primitive::kPrimInt);
break;
diff --git a/compiler/optimizing/code_generator_arm.cc b/compiler/optimizing/code_generator_arm.cc
index dd595d9..1e9fad0 100644
--- a/compiler/optimizing/code_generator_arm.cc
+++ b/compiler/optimizing/code_generator_arm.cc
@@ -1189,7 +1189,8 @@
case Primitive::kPrimFloat:
case Primitive::kPrimDouble:
- LOG(FATAL) << "Not yet implemented neg type " << neg->GetResultType();
+ locations->SetInAt(0, Location::RequiresFpuRegister());
+ locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
break;
default:
@@ -1229,8 +1230,14 @@
break;
case Primitive::kPrimFloat:
+ DCHECK(in.IsFpuRegister());
+ __ vnegs(out.As<SRegister>(), in.As<SRegister>());
+ break;
+
case Primitive::kPrimDouble:
- LOG(FATAL) << "Not yet implemented neg type " << neg->GetResultType();
+ DCHECK(in.IsFpuRegisterPair());
+ __ vnegd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()),
+ FromLowSToD(in.AsFpuRegisterPairLow<SRegister>()));
break;
default:
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index b2d9187..42e6072 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -1116,7 +1116,10 @@
case Primitive::kPrimFloat:
case Primitive::kPrimDouble:
- LOG(FATAL) << "Not yet implemented neg type " << neg->GetResultType();
+ locations->SetInAt(0, Location::RequiresFpuRegister());
+ // Output overlaps as we need a fresh (zero-initialized)
+ // register to perform subtraction from zero.
+ locations->SetOut(Location::RequiresFpuRegister());
break;
default:
@@ -1131,11 +1134,13 @@
switch (neg->GetResultType()) {
case Primitive::kPrimInt:
DCHECK(in.IsRegister());
+ DCHECK(in.Equals(out));
__ negl(out.As<Register>());
break;
case Primitive::kPrimLong:
DCHECK(in.IsRegisterPair());
+ DCHECK(in.Equals(out));
__ negl(out.AsRegisterPairLow<Register>());
// Negation is similar to subtraction from zero. The least
// significant byte triggers a borrow when it is different from
@@ -1147,8 +1152,19 @@
break;
case Primitive::kPrimFloat:
+ DCHECK(!in.Equals(out));
+ // out = 0
+ __ xorps(out.As<XmmRegister>(), out.As<XmmRegister>());
+ // out = out - in
+ __ subss(out.As<XmmRegister>(), in.As<XmmRegister>());
+ break;
+
case Primitive::kPrimDouble:
- LOG(FATAL) << "Not yet implemented neg type " << neg->GetResultType();
+ DCHECK(!in.Equals(out));
+ // out = 0
+ __ xorpd(out.As<XmmRegister>(), out.As<XmmRegister>());
+ // out = out - in
+ __ subsd(out.As<XmmRegister>(), in.As<XmmRegister>());
break;
default:
diff --git a/compiler/optimizing/code_generator_x86_64.cc b/compiler/optimizing/code_generator_x86_64.cc
index 2bd76c1..4401b9a 100644
--- a/compiler/optimizing/code_generator_x86_64.cc
+++ b/compiler/optimizing/code_generator_x86_64.cc
@@ -1102,7 +1102,10 @@
case Primitive::kPrimFloat:
case Primitive::kPrimDouble:
- LOG(FATAL) << "Not yet implemented neg type " << neg->GetResultType();
+ locations->SetInAt(0, Location::RequiresFpuRegister());
+ // Output overlaps as we need a fresh (zero-initialized)
+ // register to perform subtraction from zero.
+ locations->SetOut(Location::RequiresFpuRegister());
break;
default:
@@ -1117,17 +1120,49 @@
switch (neg->GetResultType()) {
case Primitive::kPrimInt:
DCHECK(in.IsRegister());
+ DCHECK(in.Equals(out));
__ negl(out.As<CpuRegister>());
break;
case Primitive::kPrimLong:
DCHECK(in.IsRegister());
+ DCHECK(in.Equals(out));
__ negq(out.As<CpuRegister>());
break;
case Primitive::kPrimFloat:
+ DCHECK(in.IsFpuRegister());
+ DCHECK(out.IsFpuRegister());
+ DCHECK(!in.Equals(out));
+ // TODO: Instead of computing negation as a subtraction from
+ // zero, implement it with an exclusive or with value 0x80000000
+ // (mask for bit 31, representing the sign of a single-precision
+ // floating-point number), fetched from a constant pool:
+ //
+ // xorps out, [RIP:...] // value at RIP is 0x80 00 00 00
+
+ // out = 0
+ __ xorps(out.As<XmmRegister>(), out.As<XmmRegister>());
+ // out = out - in
+ __ subss(out.As<XmmRegister>(), in.As<XmmRegister>());
+ break;
+
case Primitive::kPrimDouble:
- LOG(FATAL) << "Not yet implemented neg type " << neg->GetResultType();
+ DCHECK(in.IsFpuRegister());
+ DCHECK(out.IsFpuRegister());
+ DCHECK(!in.Equals(out));
+ // TODO: Instead of computing negation as a subtraction from
+ // zero, implement it with an exclusive or with value
+ // 0x8000000000000000 (mask for bit 63, representing the sign of
+ // a double-precision floating-point number), fetched from a
+ // constant pool:
+ //
+ // xorpd out, [RIP:...] // value at RIP is 0x80 00 00 00 00 00 00 00
+
+ // out = 0
+ __ xorpd(out.As<XmmRegister>(), out.As<XmmRegister>());
+ // out = out - in
+ __ subsd(out.As<XmmRegister>(), in.As<XmmRegister>());
break;
default: