Bug 913282: More Float32 operators: Sqrt; p=dougc,bbouvier, r=jonco,nbp

dougc for the ARM parts, bbouvier for the rest

js/src/assembler/assembler/X86Assembler.h

@@ -301,6 +301,7 @@ private:
         OP2_DIVSD_VsdWsd    = 0x5E,
         OP2_MAXSD_VsdWsd    = 0x5F,
         OP2_SQRTSD_VsdWsd   = 0x51,
+        OP2_SQRTSS_VssWss   = 0x51,
         OP2_ANDPD_VpdWpd    = 0x54,
         OP2_ORPD_VpdWpd     = 0x56,
         OP2_XORPD_VpdWpd    = 0x57,
@@ -2906,6 +2907,14 @@ public:
         m_formatter.twoByteOp(OP2_SQRTSD_VsdWsd, (RegisterID)dst, (RegisterID)src);
     }
 
+    void sqrtss_rr(XMMRegisterID src, XMMRegisterID dst)
+    {
+        spew("sqrtss     %s, %s",
+             nameFPReg(src), nameFPReg(dst));
+        m_formatter.prefix(PRE_SSE_F3);
+        m_formatter.twoByteOp(OP2_SQRTSS_VssWss, (RegisterID)dst, (RegisterID)src);
+    }
+
     void roundsd_rr(XMMRegisterID src, XMMRegisterID dst, RoundingMode mode)
     {
         spew("roundsd    %s, %s, %d",

js/src/jit/LIR-Common.h

@@ -2303,6 +2303,16 @@ class LSqrtD : public LInstructionHelper<1, 1, 0>
     }
 };
 
+// Square root of a float32.
+class LSqrtF : public LInstructionHelper<1, 1, 0>
+{
+  public:
+    LIR_HEADER(SqrtF)
+    LSqrtF(const LAllocation &num) {
+        setOperand(0, num);
+    }
+};
+
 class LAtan2D : public LCallInstructionHelper<1, 2, 1>
 {
   public:

js/src/jit/LOpcodes.h

@@ -102,6 +102,7 @@
     _(AbsI)                         \
     _(AbsD)                         \
     _(SqrtD)                        \
+    _(SqrtF)                        \
     _(Atan2D)                       \
     _(PowI)                         \
     _(PowD)                         \

js/src/jit/Lowering.cpp

@@ -1185,8 +1185,13 @@ bool
 LIRGenerator::visitSqrt(MSqrt *ins)
 {
     MDefinition *num = ins->num();
-    JS_ASSERT(num->type() == MIRType_Double);
-    LSqrtD *lir = new LSqrtD(useRegisterAtStart(num));
+    JS_ASSERT(IsFloatingPointType(num->type()));
+    if (num->type() == MIRType_Double) {
+        LSqrtD *lir = new LSqrtD(useRegisterAtStart(num));
+        return define(lir, ins);
+    }
+
+    LSqrtF *lir = new LSqrtF(useRegisterAtStart(num));
     return define(lir, ins);
 }
 

js/src/jit/MIR.cpp

@@ -2729,6 +2729,18 @@ MAsmJSCall::New(Callee callee, const Args &args, MIRType resultType, size_t spIn
     return call;
 }
 
+void
+MSqrt::trySpecializeFloat32() {
+    if (!input()->canProduceFloat32() || !CheckUsesAreFloat32Consumers(this)) {
+        if (input()->type() == MIRType_Float32)
+            ConvertDefinitionToDouble<0>(input(), this);
+        return;
+    }
+
+    setResultType(MIRType_Float32);
+    setPolicyType(MIRType_Float32);
+}
+
 bool
 jit::ElementAccessIsDenseNative(MDefinition *obj, MDefinition *id)
 {

js/src/jit/MIR.h

@@ -3455,23 +3455,24 @@ class MAbs
 // Inline implementation of Math.sqrt().
 class MSqrt
   : public MUnaryInstruction,
-    public DoublePolicy<0>
+    public FloatingPointPolicy<0>
 {
-    MSqrt(MDefinition *num)
+    MSqrt(MDefinition *num, MIRType type)
       : MUnaryInstruction(num)
     {
-        setResultType(MIRType_Double);
+        setResultType(type);
+        setPolicyType(type);
         setMovable();
     }
 
   public:
     INSTRUCTION_HEADER(Sqrt)
     static MSqrt *New(MDefinition *num) {
-        return new MSqrt(num);
+        return new MSqrt(num, MIRType_Double);
     }
     static MSqrt *NewAsmJS(MDefinition *num, MIRType type) {
-        JS_ASSERT(type == MIRType_Double);
-        return new MSqrt(num);
+        JS_ASSERT(IsFloatingPointType(type));
+        return new MSqrt(num, type);
     }
     MDefinition *num() const {
         return getOperand(0);
@@ -3487,6 +3488,9 @@ class MSqrt
         return AliasSet::None();
     }
     void computeRange();
+
+    bool isFloat32Commutative() const { return true; }
+    void trySpecializeFloat32();
 };
 
 // Inline implementation of atan2 (arctangent of y/x).

js/src/jit/arm/CodeGenerator-arm.cpp

@@ -348,6 +348,15 @@ CodeGeneratorARM::visitSqrtD(LSqrtD *ins)
     return true;
 }
 
+bool
+CodeGeneratorARM::visitSqrtF(LSqrtF *ins)
+{
+    FloatRegister input = ToFloatRegister(ins->input());
+    FloatRegister output = ToFloatRegister(ins->output());
+    masm.ma_vsqrt_f32(input, output);
+    return true;
+}
+
 bool
 CodeGeneratorARM::visitAddI(LAddI *ins)
 {

js/src/jit/arm/CodeGenerator-arm.h

@@ -67,6 +67,7 @@ class CodeGeneratorARM : public CodeGeneratorShared
     virtual bool visitMinMaxD(LMinMaxD *ins);
     virtual bool visitAbsD(LAbsD *ins);
     virtual bool visitSqrtD(LSqrtD *ins);
+    virtual bool visitSqrtF(LSqrtF *ins);
     virtual bool visitAddI(LAddI *ins);
     virtual bool visitSubI(LSubI *ins);
     virtual bool visitBitNotI(LBitNotI *ins);

js/src/jit/arm/MacroAssembler-arm.cpp

@@ -1431,6 +1431,12 @@ MacroAssemblerARM::ma_vsqrt(FloatRegister src, FloatRegister dest, Condition cc)
     as_vsqrt(dest, src, cc);
 }
 
+void
+MacroAssemblerARM::ma_vsqrt_f32(FloatRegister src, FloatRegister dest, Condition cc)
+{
+    as_vsqrt(VFPRegister(dest).singleOverlay(), VFPRegister(src).singleOverlay(), cc);
+}
+
 union DoublePun
 {
     struct

js/src/jit/arm/MacroAssembler-arm.h

@@ -319,6 +319,7 @@ class MacroAssemblerARM : public Assembler
     void ma_vabs(FloatRegister src, FloatRegister dest, Condition cc = Always);
 
     void ma_vsqrt(FloatRegister src, FloatRegister dest, Condition cc = Always);
+    void ma_vsqrt_f32(FloatRegister src, FloatRegister dest, Condition cc = Always);
 
     void ma_vimm(double value, FloatRegister dest, Condition cc = Always);
     void ma_vimm_f32(float value, FloatRegister dest, Condition cc = Always);

js/src/jit/shared/Assembler-x86-shared.h

@@ -1452,6 +1452,10 @@ class AssemblerX86Shared
         JS_ASSERT(HasSSE2());
         masm.sqrtsd_rr(src.code(), dest.code());
     }
+    void sqrtss(const FloatRegister &src, const FloatRegister &dest) {
+        JS_ASSERT(HasSSE2());
+        masm.sqrtss_rr(src.code(), dest.code());
+    }
     void roundsd(const FloatRegister &src, const FloatRegister &dest,
                  JSC::X86Assembler::RoundingMode mode)
     {

js/src/jit/shared/CodeGenerator-x86-shared.cpp

@@ -482,6 +482,15 @@ CodeGeneratorX86Shared::visitSqrtD(LSqrtD *ins)
     return true;
 }
 
+bool
+CodeGeneratorX86Shared::visitSqrtF(LSqrtF *ins)
+{
+    FloatRegister input = ToFloatRegister(ins->input());
+    FloatRegister output = ToFloatRegister(ins->output());
+    masm.sqrtss(input, output);
+    return true;
+}
+
 bool
 CodeGeneratorX86Shared::visitPowHalfD(LPowHalfD *ins)
 {

js/src/jit/shared/CodeGenerator-x86-shared.h

@@ -83,6 +83,7 @@ class CodeGeneratorX86Shared : public CodeGeneratorShared
     virtual bool visitMinMaxD(LMinMaxD *ins);
     virtual bool visitAbsD(LAbsD *ins);
     virtual bool visitSqrtD(LSqrtD *ins);
+    virtual bool visitSqrtF(LSqrtF *ins);
     virtual bool visitPowHalfD(LPowHalfD *ins);
     virtual bool visitAddI(LAddI *ins);
     virtual bool visitSubI(LSubI *ins);