--- a/js/src/jit-test/tests/asm.js/testSIMD.js
+++ b/js/src/jit-test/tests/asm.js/testSIMD.js
@@ -510,64 +510,70 @@ CheckI4(WWI, 'var x = i4(1,2,3,4); x = w
 // True yields all bits set to 1 (i.e as an int32, 0xFFFFFFFF === -1), false
 // yields all bits set to 0 (i.e 0).
 const T = -1;
 const F = 0;
 assertAsmTypeFail('glob', USE_ASM + I32 + "var lt=i4.lessThanOrEqual; function f() {} return f");
 assertAsmTypeFail('glob', USE_ASM + I32 + "var ge=i4.greaterThanOrEqual; function f() {} return f");
 assertAsmTypeFail('glob', USE_ASM + I32 + "var ne=i4.notEqual; function f() {} return f");
 
-const LTI32 = 'var lt = i4.lessThan';
-const GTI32 = 'var gt = i4.greaterThan';
-const EQI32 = 'var eq = i4.equal';
+const LTI32 = 'var lt = i4.lessThan;';
+const GTI32 = 'var gt = i4.greaterThan;';
+const EQI32 = 'var eq = i4.equal;';
 
 CheckI4(LTI32, 'var x=i4(1,2,3,4); var y=i4(-1,1,0,2); x=lt(x,y)', [F, F, F, F]);
 CheckI4(LTI32, 'var x=i4(-1,1,0,2); var y=i4(1,2,3,4); x=lt(x,y)', [T, T, T, T]);
 CheckI4(LTI32, 'var x=i4(1,0,3,4); var y=i4(1,1,7,0); x=lt(x,y)', [F, T, T, F]);
 
 CheckI4(EQI32, 'var x=i4(1,2,3,4); var y=i4(-1,1,0,2); x=eq(x,y)', [F, F, F, F]);
 CheckI4(EQI32, 'var x=i4(-1,1,0,2); var y=i4(1,2,3,4); x=eq(x,y)', [F, F, F, F]);
 CheckI4(EQI32, 'var x=i4(1,0,3,4); var y=i4(1,1,7,0); x=eq(x,y)', [T, F, F, F]);
 
 CheckI4(GTI32, 'var x=i4(1,2,3,4); var y=i4(-1,1,0,2); x=gt(x,y)', [T, T, T, T]);
 CheckI4(GTI32, 'var x=i4(-1,1,0,2); var y=i4(1,2,3,4); x=gt(x,y)', [F, F, F, F]);
 CheckI4(GTI32, 'var x=i4(1,0,3,4); var y=i4(1,1,7,0); x=gt(x,y)', [F, F, F, T]);
 
-const LTF32 = 'var lt=f4.lessThan';
-const LEF32 = 'var le=f4.lessThanOrEqual';
-const GTF32 = 'var gt=f4.greaterThan';
-const GEF32 = 'var ge=f4.greaterThanOrEqual';
-const EQF32 = 'var eq=f4.equal';
-const NEF32 = 'var ne=f4.notEqual';
+const LTF32 = 'var lt=f4.lessThan;';
+const LEF32 = 'var le=f4.lessThanOrEqual;';
+const GTF32 = 'var gt=f4.greaterThan;';
+const GEF32 = 'var ge=f4.greaterThanOrEqual;';
+const EQF32 = 'var eq=f4.equal;';
+const NEF32 = 'var ne=f4.notEqual;';
 
 assertAsmTypeFail('glob', USE_ASM + F32 + "var lt=f4.lessThan; function f() {var x=f4(1,2,3,4); var y=f4(5,6,7,8); x=lt(x,y);} return f");
 
 CheckF4Comp(LTF32, 'var y=f4(1,2,3,4);  var z=f4(-1,1,0,2); var x=i4(0,0,0,0); x=lt(y,z)', [F, F, F, F]);
 CheckF4Comp(LTF32, 'var y=f4(-1,1,0,2); var z=f4(1,2,3,4);  var x=i4(0,0,0,0); x=lt(y,z)', [T, T, T, T]);
 CheckF4Comp(LTF32, 'var y=f4(1,0,3,4);  var z=f4(1,1,7,0);  var x=i4(0,0,0,0); x=lt(y,z)', [F, T, T, F]);
+CheckF4Comp(LTF32 + 'const nan = glob.NaN; const fround=glob.Math.fround', 'var y=f4(0,0,0,0); var z=f4(0,0,0,0); var x=i4(0,0,0,0); y=f4(fround(0.0),fround(-0.0),fround(0.0),fround(nan)); z=f4(fround(-0.0),fround(0.0),fround(nan),fround(0.0)); x=lt(y,z);', [F, F, F, F]);
 
 CheckF4Comp(LEF32, 'var y=f4(1,2,3,4);  var z=f4(-1,1,0,2); var x=i4(0,0,0,0); x=le(y,z)', [F, F, F, F]);
 CheckF4Comp(LEF32, 'var y=f4(-1,1,0,2); var z=f4(1,2,3,4);  var x=i4(0,0,0,0); x=le(y,z)', [T, T, T, T]);
 CheckF4Comp(LEF32, 'var y=f4(1,0,3,4);  var z=f4(1,1,7,0);  var x=i4(0,0,0,0); x=le(y,z)', [T, T, T, F]);
+CheckF4Comp(LEF32 + 'const nan = glob.NaN; const fround=glob.Math.fround', 'var y=f4(0,0,0,0); var z=f4(0,0,0,0); var x=i4(0,0,0,0); y=f4(fround(0.0),fround(-0.0),fround(0.0),fround(nan)); z=f4(fround(-0.0),fround(0.0),fround(nan),fround(0.0)); x=le(y,z);', [T, T, F, F]);
 
 CheckF4Comp(EQF32, 'var y=f4(1,2,3,4);  var z=f4(-1,1,0,2); var x=i4(0,0,0,0); x=eq(y,z)', [F, F, F, F]);
 CheckF4Comp(EQF32, 'var y=f4(-1,1,0,2); var z=f4(1,2,3,4);  var x=i4(0,0,0,0); x=eq(y,z)', [F, F, F, F]);
 CheckF4Comp(EQF32, 'var y=f4(1,0,3,4);  var z=f4(1,1,7,0);  var x=i4(0,0,0,0); x=eq(y,z)', [T, F, F, F]);
+CheckF4Comp(EQF32 + 'const nan = glob.NaN; const fround=glob.Math.fround', 'var y=f4(0,0,0,0); var z=f4(0,0,0,0); var x=i4(0,0,0,0); y=f4(fround(0.0),fround(-0.0),fround(0.0),fround(nan)); z=f4(fround(-0.0),fround(0.0),fround(nan),fround(0.0)); x=eq(y,z);', [T, T, F, F]);
 
 CheckF4Comp(NEF32, 'var y=f4(1,2,3,4);  var z=f4(-1,1,0,2); var x=i4(0,0,0,0); x=ne(y,z)', [T, T, T, T]);
 CheckF4Comp(NEF32, 'var y=f4(-1,1,0,2); var z=f4(1,2,3,4);  var x=i4(0,0,0,0); x=ne(y,z)', [T, T, T, T]);
 CheckF4Comp(NEF32, 'var y=f4(1,0,3,4);  var z=f4(1,1,7,0);  var x=i4(0,0,0,0); x=ne(y,z)', [F, T, T, T]);
+CheckF4Comp(NEF32 + 'const nan = glob.NaN; const fround=glob.Math.fround', 'var y=f4(0,0,0,0); var z=f4(0,0,0,0); var x=i4(0,0,0,0); y=f4(fround(0.0),fround(-0.0),fround(0.0),fround(nan)); z=f4(fround(-0.0),fround(0.0),fround(nan),fround(0.0)); x=ne(y,z);', [F, F, T, T]);
 
 CheckF4Comp(GTF32, 'var y=f4(1,2,3,4);  var z=f4(-1,1,0,2); var x=i4(0,0,0,0); x=gt(y,z)', [T, T, T, T]);
 CheckF4Comp(GTF32, 'var y=f4(-1,1,0,2); var z=f4(1,2,3,4);  var x=i4(0,0,0,0); x=gt(y,z)', [F, F, F, F]);
 CheckF4Comp(GTF32, 'var y=f4(1,0,3,4);  var z=f4(1,1,7,0);  var x=i4(0,0,0,0); x=gt(y,z)', [F, F, F, T]);
+CheckF4Comp(GTF32 + 'const nan = glob.NaN; const fround=glob.Math.fround', 'var y=f4(0,0,0,0); var z=f4(0,0,0,0); var x=i4(0,0,0,0); y=f4(fround(0.0),fround(-0.0),fround(0.0),fround(nan)); z=f4(fround(-0.0),fround(0.0),fround(nan),fround(0.0)); x=gt(y,z);', [F, F, F, F]);
 
 CheckF4Comp(GEF32, 'var y=f4(1,2,3,4);  var z=f4(-1,1,0,2); var x=i4(0,0,0,0); x=ge(y,z)', [T, T, T, T]);
 CheckF4Comp(GEF32, 'var y=f4(-1,1,0,2); var z=f4(1,2,3,4);  var x=i4(0,0,0,0); x=ge(y,z)', [F, F, F, F]);
 CheckF4Comp(GEF32, 'var y=f4(1,0,3,4);  var z=f4(1,1,7,0);  var x=i4(0,0,0,0); x=ge(y,z)', [T, F, F, T]);
+CheckF4Comp(GEF32 + 'const nan = glob.NaN; const fround=glob.Math.fround', 'var y=f4(0,0,0,0); var z=f4(0,0,0,0); var x=i4(0,0,0,0); y=f4(fround(0.0),fround(-0.0),fround(0.0),fround(nan)); z=f4(fround(-0.0),fround(0.0),fround(nan),fround(0.0)); x=ge(y,z);', [T, T, F, F]);
 
 // Conversions operators
 const CVTIF = 'var cvt=f4.fromInt32x4;';
 const CVTFI = 'var cvt=i4.fromFloat32x4;';
 
 assertAsmTypeFail('glob', USE_ASM + I32 + "var cvt=i4.fromInt32x4; return {}");
 assertAsmTypeFail('glob', USE_ASM + F32 + "var cvt=f4.fromFloat32x4; return {}");
 assertAsmTypeFail('glob', USE_ASM + I32 + F32 + CVTIF + "function f() {var x=i4(1,2,3,4); x=cvt(x);} return f");

--- a/js/src/jit/Lowering.cpp
+++ b/js/src/jit/Lowering.cpp
@@ -3803,22 +3803,22 @@ LIRGenerator::visitSimdSignMask(MSimdSig
 
 bool
 LIRGenerator::visitSimdBinaryComp(MSimdBinaryComp *ins)
 {
     MOZ_ASSERT(ins->type() == MIRType_Int32x4);
 
     if (ins->compareType() == MSimdBinaryComp::CompareInt32x4) {
         LSimdBinaryCompIx4 *add = new(alloc()) LSimdBinaryCompIx4();
-        return lowerForFPU(add, ins, ins->lhs(), ins->rhs());
+        return lowerForCompIx4(add, ins, ins->lhs(), ins->rhs());
     }
 
     if (ins->compareType() == MSimdBinaryComp::CompareFloat32x4) {
         LSimdBinaryCompFx4 *add = new(alloc()) LSimdBinaryCompFx4();
-        return lowerForFPU(add, ins, ins->lhs(), ins->rhs());
+        return lowerForCompFx4(add, ins, ins->lhs(), ins->rhs());
     }
 
     MOZ_CRASH("Unknown compare type when comparing values");
     return false;
 }
 
 bool
 LIRGenerator::visitSimdBinaryArith(MSimdBinaryArith *ins)

--- a/js/src/jit/MIR.h
+++ b/js/src/jit/MIR.h
@@ -1606,16 +1606,31 @@ class MSimdBinaryComp : public MBinaryIn
 
     AliasSet getAliasSet() const {
         return AliasSet::None();
     }
 
     Operation operation() const { return operation_; }
     CompareType compareType() const { return compareType_; }
 
+    // Swap the operands and reverse the comparison predicate.
+    void reverse() {
+        switch (operation()) {
+          case greaterThan:        operation_ = lessThan; break;
+          case greaterThanOrEqual: operation_ = lessThanOrEqual; break;
+          case lessThan:           operation_ = greaterThan; break;
+          case lessThanOrEqual:    operation_ = greaterThanOrEqual; break;
+          case equal:
+          case notEqual:
+            break;
+          default: MOZ_CRASH("Unexpected compare operation");
+        }
+        swapOperands();
+    }
+
     bool congruentTo(const MDefinition *ins) const {
         if (!binaryCongruentTo(ins))
             return false;
         return operation_ == ins->toSimdBinaryComp()->operation();
     }
 
     ALLOW_CLONE(MSimdBinaryComp)
 };

--- a/js/src/jit/arm/Lowering-arm.h
+++ b/js/src/jit/arm/Lowering-arm.h
@@ -53,16 +53,28 @@ class LIRGeneratorARM : public LIRGenera
                      MDefinition *input);
     bool lowerForALU(LInstructionHelper<1, 2, 0> *ins, MDefinition *mir,
                      MDefinition *lhs, MDefinition *rhs);
 
     bool lowerForFPU(LInstructionHelper<1, 1, 0> *ins, MDefinition *mir,
                      MDefinition *src);
     bool lowerForFPU(LInstructionHelper<1, 2, 0> *ins, MDefinition *mir,
                      MDefinition *lhs, MDefinition *rhs);
+
+    bool lowerForCompIx4(LSimdBinaryCompIx4 *ins, MSimdBinaryComp *mir,
+                         MDefinition *lhs, MDefinition *rhs)
+    {
+        return lowerForFPU(ins, mir, lhs, rhs);
+    }
+    bool lowerForCompFx4(LSimdBinaryCompFx4 *ins, MSimdBinaryComp *mir,
+                         MDefinition *lhs, MDefinition *rhs)
+    {
+        return lowerForFPU(ins, mir, lhs, rhs);
+    }
+
     bool lowerForBitAndAndBranch(LBitAndAndBranch *baab, MInstruction *mir,
                                  MDefinition *lhs, MDefinition *rhs);
     bool lowerConstantDouble(double d, MInstruction *ins);
     bool lowerConstantFloat32(float d, MInstruction *ins);
     bool lowerTruncateDToInt32(MTruncateToInt32 *ins);
     bool lowerTruncateFToInt32(MTruncateToInt32 *ins);
     bool lowerDivI(MDiv *div);
     bool lowerModI(MMod *mod);

--- a/js/src/jit/mips/Lowering-mips.h
+++ b/js/src/jit/mips/Lowering-mips.h
@@ -53,16 +53,28 @@ class LIRGeneratorMIPS : public LIRGener
                      MDefinition *input);
     bool lowerForALU(LInstructionHelper<1, 2, 0> *ins, MDefinition *mir,
                      MDefinition *lhs, MDefinition *rhs);
 
     bool lowerForFPU(LInstructionHelper<1, 1, 0> *ins, MDefinition *mir,
                      MDefinition *src);
     bool lowerForFPU(LInstructionHelper<1, 2, 0> *ins, MDefinition *mir,
                      MDefinition *lhs, MDefinition *rhs);
+
+    bool lowerForCompIx4(LSimdBinaryCompIx4 *ins, MSimdBinaryComp *mir,
+                         MDefinition *lhs, MDefinition *rhs)
+    {
+        return lowerForFPU(ins, mir, lhs, rhs);
+    }
+    bool lowerForCompFx4(LSimdBinaryCompFx4 *ins, MSimdBinaryComp *mir,
+                         MDefinition *lhs, MDefinition *rhs)
+    {
+        return lowerForFPU(ins, mir, lhs, rhs);
+    }
+
     bool lowerForBitAndAndBranch(LBitAndAndBranch *baab, MInstruction *mir,
                                  MDefinition *lhs, MDefinition *rhs);
     bool lowerConstantDouble(double d, MInstruction *ins);
     bool lowerConstantFloat32(float d, MInstruction *ins);
     bool lowerTruncateDToInt32(MTruncateToInt32 *ins);
     bool lowerTruncateFToInt32(MTruncateToInt32 *ins);
     bool lowerDivI(MDiv *div);
     bool lowerModI(MMod *mod);

--- a/js/src/jit/none/Lowering-none.h
+++ b/js/src/jit/none/Lowering-none.h
@@ -37,16 +37,24 @@ class LIRGeneratorNone : public LIRGener
     bool lowerForShift(LInstructionHelper<1, 2, 0> *, MDefinition *, MDefinition *, MDefinition *) {
         MOZ_CRASH();
     }
     bool lowerUrshD(MUrsh *) { MOZ_CRASH(); }
     template <typename T>
     bool lowerForALU(T, MDefinition *, MDefinition *, MDefinition *v = nullptr) { MOZ_CRASH(); }
     template <typename T>
     bool lowerForFPU(T, MDefinition *, MDefinition *, MDefinition *v = nullptr) { MOZ_CRASH(); }
+    bool lowerForCompIx4(LSimdBinaryCompIx4 *ins, MSimdBinaryComp *mir,
+                         MDefinition *lhs, MDefinition *rhs) {
+        MOZ_CRASH();
+    }
+    bool lowerForCompFx4(LSimdBinaryCompFx4 *ins, MSimdBinaryComp *mir,
+                         MDefinition *lhs, MDefinition *rhs) {
+        MOZ_CRASH();
+    }
     bool lowerForBitAndAndBranch(LBitAndAndBranch *, MInstruction *,
                                  MDefinition *, MDefinition *) {
         MOZ_CRASH();
     }
 
     bool lowerConstantDouble(double, MInstruction *) { MOZ_CRASH(); }
     bool lowerConstantFloat32(float, MInstruction *) { MOZ_CRASH(); }
     bool lowerTruncateDToInt32(MTruncateToInt32 *) { MOZ_CRASH(); }

--- a/js/src/jit/shared/CodeGenerator-x86-shared.cpp
+++ b/js/src/jit/shared/CodeGenerator-x86-shared.cpp
@@ -2408,17 +2408,17 @@ CodeGeneratorX86Shared::visitSimdBinaryC
         if (rhs.kind() == Operand::FPREG)
             masm.moveAlignedInt32x4(ToFloatRegister(ins->rhs()), ScratchSimdReg);
         else
             masm.loadAlignedInt32x4(rhs, ScratchSimdReg);
 
         // scr := scr > lhs (i.e. lhs < rhs)
         // Improve by doing custom lowering (rhs is tied to the output register)
         masm.packedGreaterThanInt32x4(ToOperand(ins->lhs()), ScratchSimdReg);
-        masm.moveAlignedInt32x4(ScratchFloat32Reg, lhs);
+        masm.moveAlignedInt32x4(ScratchSimdReg, lhs);
         return true;
       case MSimdBinaryComp::notEqual:
       case MSimdBinaryComp::greaterThanOrEqual:
       case MSimdBinaryComp::lessThanOrEqual:
         // These operations are not part of the spec. so are not implemented.
         break;
     }
     MOZ_CRASH("unexpected SIMD op");
@@ -2441,21 +2441,20 @@ CodeGeneratorX86Shared::visitSimdBinaryC
         return true;
       case MSimdBinaryComp::lessThanOrEqual:
         masm.cmpps(rhs, lhs, 0x2);
         return true;
       case MSimdBinaryComp::notEqual:
         masm.cmpps(rhs, lhs, 0x4);
         return true;
       case MSimdBinaryComp::greaterThanOrEqual:
-        masm.cmpps(rhs, lhs, 0x5);
-        return true;
       case MSimdBinaryComp::greaterThan:
-        masm.cmpps(rhs, lhs, 0x6);
-        return true;
+        // We reverse these before register allocation so that we don't have to
+        // copy into and out of temporaries after codegen.
+        MOZ_CRASH("lowering should have reversed this");
     }
     MOZ_CRASH("unexpected SIMD op");
 }
 
 bool
 CodeGeneratorX86Shared::visitSimdBinaryArithIx4(LSimdBinaryArithIx4 *ins)
 {
     FloatRegister lhs = ToFloatRegister(ins->lhs());

--- a/js/src/jit/shared/Lowering-x86-shared.cpp
+++ b/js/src/jit/shared/Lowering-x86-shared.cpp
@@ -12,16 +12,17 @@
 
 #include "jit/shared/Lowering-shared-inl.h"
 
 using namespace js;
 using namespace js::jit;
 
 using mozilla::Abs;
 using mozilla::FloorLog2;
+using mozilla::Swap;
 
 LTableSwitch *
 LIRGeneratorX86Shared::newLTableSwitch(const LAllocation &in, const LDefinition &inputCopy,
                                        MTableSwitch *tableswitch)
 {
     return new(alloc()) LTableSwitch(in, inputCopy, temp(), tableswitch);
 }
 
@@ -103,16 +104,41 @@ bool
 LIRGeneratorX86Shared::lowerForFPU(LInstructionHelper<1, 2, 0> *ins, MDefinition *mir, MDefinition *lhs, MDefinition *rhs)
 {
     ins->setOperand(0, useRegisterAtStart(lhs));
     ins->setOperand(1, lhs != rhs ? use(rhs) : useAtStart(rhs));
     return defineReuseInput(ins, mir, 0);
 }
 
 bool
+LIRGeneratorX86Shared::lowerForCompIx4(LSimdBinaryCompIx4 *ins, MSimdBinaryComp *mir, MDefinition *lhs, MDefinition *rhs)
+{
+    return lowerForALU(ins, mir, lhs, rhs);
+}
+
+bool
+LIRGeneratorX86Shared::lowerForCompFx4(LSimdBinaryCompFx4 *ins, MSimdBinaryComp *mir, MDefinition *lhs, MDefinition *rhs)
+{
+    // Swap the operands around to fit the instructions that x86 actually has.
+    // We do this here, before register allocation, so that we don't need
+    // temporaries and copying afterwards.
+    switch (mir->operation()) {
+      case MSimdBinaryComp::greaterThan:
+      case MSimdBinaryComp::greaterThanOrEqual:
+        mir->reverse();
+        Swap(lhs, rhs);
+        break;
+      default:
+        break;
+    }
+
+    return lowerForFPU(ins, mir, lhs, rhs);
+}
+
+bool
 LIRGeneratorX86Shared::lowerForBitAndAndBranch(LBitAndAndBranch *baab, MInstruction *mir,
                                                MDefinition *lhs, MDefinition *rhs)
 {
     baab->setOperand(0, useRegisterAtStart(lhs));
     baab->setOperand(1, useRegisterOrConstantAtStart(rhs));
     return add(baab, mir);
 }
 

--- a/js/src/jit/shared/Lowering-x86-shared.h
+++ b/js/src/jit/shared/Lowering-x86-shared.h
@@ -28,16 +28,20 @@ class LIRGeneratorX86Shared : public LIR
     bool visitPowHalf(MPowHalf *ins);
     bool lowerForShift(LInstructionHelper<1, 2, 0> *ins, MDefinition *mir, MDefinition *lhs,
                        MDefinition *rhs);
     bool lowerForALU(LInstructionHelper<1, 1, 0> *ins, MDefinition *mir, MDefinition *input);
     bool lowerForALU(LInstructionHelper<1, 2, 0> *ins, MDefinition *mir, MDefinition *lhs,
                      MDefinition *rhs);
     bool lowerForFPU(LInstructionHelper<1, 2, 0> *ins, MDefinition *mir, MDefinition *lhs,
                      MDefinition *rhs);
+    bool lowerForCompIx4(LSimdBinaryCompIx4 *ins, MSimdBinaryComp *mir,
+                         MDefinition *lhs, MDefinition *rhs);
+    bool lowerForCompFx4(LSimdBinaryCompFx4 *ins, MSimdBinaryComp *mir,
+                         MDefinition *lhs, MDefinition *rhs);
     bool lowerForBitAndAndBranch(LBitAndAndBranch *baab, MInstruction *mir,
                                  MDefinition *lhs, MDefinition *rhs);
     bool visitConstant(MConstant *ins);
     bool visitAsmJSNeg(MAsmJSNeg *ins);
     bool lowerMulI(MMul *mul, MDefinition *lhs, MDefinition *rhs);
     bool lowerDivI(MDiv *div);
     bool lowerModI(MMod *mod);
     bool lowerUDiv(MDiv *div);