Add SSE2 optimisations

js/src/jit/IonOptimizationLevels.cpp

@@ -95,12 +95,22 @@ OptimizationInfo::compilerWarmUpThreshold(JSScript* script, jsbytecode* pc) cons
     // threshold to improve the compilation's type information and hopefully
     // avoid later recompilation.
 
-    if (script->length() > MAX_MAIN_THREAD_SCRIPT_SIZE)
-        warmUpThreshold *= (script->length() / (double) MAX_MAIN_THREAD_SCRIPT_SIZE);
+    if (script->length() > MAX_MAIN_THREAD_SCRIPT_SIZE) {
+        // Avoid pathological thresholds on very large scripts: large warm-up
+        // counts delay optimization too much for hot UI/update code.
+        double ratio = script->length() / (double) MAX_MAIN_THREAD_SCRIPT_SIZE;
+        if (ratio > 4.0)
+            ratio = 4.0;
+        warmUpThreshold *= ratio;
+    }
 
     uint32_t numLocalsAndArgs = NumLocalsAndArgs(script);
-    if (numLocalsAndArgs > MAX_MAIN_THREAD_LOCALS_AND_ARGS)
-        warmUpThreshold *= (numLocalsAndArgs / (double) MAX_MAIN_THREAD_LOCALS_AND_ARGS);
+    if (numLocalsAndArgs > MAX_MAIN_THREAD_LOCALS_AND_ARGS) {
+        double ratio = numLocalsAndArgs / (double) MAX_MAIN_THREAD_LOCALS_AND_ARGS;
+        if (ratio > 4.0)
+            ratio = 4.0;
+        warmUpThreshold *= ratio;
+    }
 
     if (!pc || JitOptions.eagerCompilation)
         return warmUpThreshold;
@@ -110,7 +120,21 @@ OptimizationInfo::compilerWarmUpThreshold(JSScript* script, jsbytecode* pc) cons
     // Note that the loop depth is always > 0 so we will prefer non-OSR over OSR.
     uint32_t loopDepth = LoopEntryDepthHint(pc);
     MOZ_ASSERT(loopDepth > 0);
-    return warmUpThreshold + loopDepth * 100;
+
+    // jQuery-style code often executes many small hot loops. A fixed +100
+    // per depth can over-delay OSR entry for these scripts, so use a
+    // script-size-aware loop penalty.
+    uint32_t perDepthPenalty;
+    if (JitOptions.isSmallFunction(script)) {
+        perDepthPenalty = 25;
+    } else {
+        perDepthPenalty = warmUpThreshold / 8;
+        if (perDepthPenalty < 50)
+            perDepthPenalty = 50;
+        if (perDepthPenalty > 200)
+            perDepthPenalty = 200;
+    }
+    return warmUpThreshold + loopDepth * perDepthPenalty;
 }
 
 OptimizationLevelInfo::OptimizationLevelInfo()

js/src/jit/x86/Lowering-x86.cpp

@@ -631,8 +631,7 @@ LIRGeneratorX86::visitInt64ToFloatingPoint(MInt64ToFloatingPoint* ins)
     MOZ_ASSERT(opd->type() == MIRType::Int64);
     MOZ_ASSERT(IsFloatingPointType(ins->type()));
 
-    LDefinition maybeTemp =
-        (ins->isUnsigned() && AssemblerX86Shared::HasSSE3()) ? temp() : LDefinition::BogusTemp();
+    LDefinition maybeTemp = LDefinition::BogusTemp();
 
     define(new(alloc()) LInt64ToFloatingPoint(useInt64Register(opd), maybeTemp), ins);
 }

js/src/jit/x86/MacroAssembler-x86.cpp

@@ -25,36 +25,14 @@ static const double TO_DOUBLE_HIGH_SCALE = 0x100000000;
 bool
 MacroAssemblerX86::convertUInt64ToDoubleNeedsTemp()
 {
-    return HasSSE3();
+    return false;
 }
 
 void
 MacroAssemblerX86::convertUInt64ToDouble(Register64 src, FloatRegister dest, Register temp)
 {
-    // SUBPD needs SSE2, HADDPD needs SSE3.
-    if (!HasSSE3()) {
-        MOZ_ASSERT(temp == Register::Invalid());
-
-        // Zero the dest register to break dependencies, see convertInt32ToDouble.
-        zeroDouble(dest);
-
-        asMasm().Push(src.high);
-        asMasm().Push(src.low);
-        fild(Operand(esp, 0));
-
-        Label notNegative;
-        asMasm().branch32(Assembler::NotSigned, src.high, Imm32(0), &notNegative);
-        double add_constant = 18446744073709551616.0; // 2^64
-        store64(Imm64(mozilla::BitwiseCast<uint64_t>(add_constant)), Address(esp, 0));
-        fld(Operand(esp, 0));
-        faddp();
-        bind(&notNegative);
-
-        fstp(Operand(esp, 0));
-        vmovsd(Address(esp, 0), dest);
-        asMasm().freeStack(2*sizeof(intptr_t));
-        return;
-    }
+    (void) temp;
+    MOZ_ASSERT(HasSSE2());
 
     // Following operation uses entire 128-bit of dest XMM register.
     // Currently higher 64-bit is free when we have access to lower 64-bit.
@@ -113,7 +91,8 @@ MacroAssemblerX86::convertUInt64ToDouble(Register64 src, FloatRegister dest, Reg
     //   LO(dest) = double(0x HHHHHHHH 00000000) + double(0x 00000000 LLLLLLLL)
     //            = double(0x HHHHHHHH LLLLLLLL)
     //            = double(src)
-    vhaddpd(dest128, dest128);
+    vmovhlps(dest128, dest128, ScratchSimd128Reg);
+    vaddsd(ScratchSimd128Reg, dest128, dest128);
 }
 
 void