1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/public/jit/x86/assembler-x86.cpp	Thu Aug 08 20:26:36 2013 -0700
     1.3 @@ -0,0 +1,33 @@
     1.4 +/**
     1.5 + * vim: set ts=8 sts=2 sw=2 tw=99 et:
     1.6 + * =============================================================================
     1.7 + * SourcePawn JIT SDK
     1.8 + * Copyright (C) 2004-2008 AlliedModders LLC.  All rights reserved.
     1.9 + * =============================================================================
    1.10 + *
    1.11 + * This program is free software; you can redistribute it and/or modify it under
    1.12 + * the terms of the GNU General Public License, version 3.0, as published by the
    1.13 + * Free Software Foundation.
    1.14 + * 
    1.15 + * This program is distributed in the hope that it will be useful, but WITHOUT
    1.16 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
    1.17 + * FOR A PARTICULAR PURPOSE.  See the GNU General Public License for more
    1.18 + * details.
    1.19 + *
    1.20 + * You should have received a copy of the GNU General Public License along with
    1.21 + * this program.  If not, see <http://www.gnu.org/licenses/>.
    1.22 + *
    1.23 + * As a special exception, AlliedModders LLC gives you permission to link the
    1.24 + * code of this program (as well as its derivative works) to "Half-Life 2," the
    1.25 + * "Source Engine," the "SourcePawn JIT," and any Game MODs that run on software
    1.26 + * by the Valve Corporation.  You must obey the GNU General Public License in
    1.27 + * all respects for all other code used.  Additionally, AlliedModders LLC grants
    1.28 + * this exception to all derivative works.  AlliedModders LLC defines further
    1.29 + * exceptions, found in LICENSE.txt (as of this writing, version JULY-31-2007),
    1.30 + * or <http://www.sourcemod.net/license.php>.
    1.31 + *
    1.32 + * Version: $Id$
    1.33 + */
    1.34 +#include <assembler-x86.h>
    1.35 +
    1.36 +CPUFeatures AssemblerX86::X86Features;

     2.1 --- a/public/jit/x86/assembler-x86.h	Thu Aug 08 20:21:37 2013 -0700
     2.2 +++ b/public/jit/x86/assembler-x86.h	Thu Aug 08 20:26:36 2013 -0700
     2.3 @@ -94,6 +94,25 @@
     2.4    }
     2.5  };
     2.6  
     2.7 +struct CPUFeatures
     2.8 +{
     2.9 +  CPUFeatures()
    2.10 +  {
    2.11 +    memset(this, 0, sizeof(*this));
    2.12 +  }
    2.13 +
    2.14 +  bool fpu;
    2.15 +  bool mmx;
    2.16 +  bool sse;
    2.17 +  bool sse2;
    2.18 +  bool sse3;
    2.19 +  bool ssse3;
    2.20 +  bool sse4_1;
    2.21 +  bool sse4_2;
    2.22 +  bool avx;
    2.23 +  bool avx2;
    2.24 +};
    2.25 +
    2.26  const Register eax = { 0 };
    2.27  const Register ecx = { 1 };
    2.28  const Register edx = { 2 };
    2.29 @@ -299,7 +318,19 @@
    2.30  
    2.31  class AssemblerX86 : public Assembler
    2.32  {
    2.33 + private:
    2.34 +  // List of processor features; to be used, this must be filled in at
    2.35 +  // startup.
    2.36 +  static CPUFeatures X86Features; 
    2.37 +
    2.38   public:
    2.39 +  static void SetFeatures(const CPUFeatures &features) {
    2.40 +    X86Features = features;
    2.41 +  }
    2.42 +  static const CPUFeatures &Features() {
    2.43 +    return X86Features;
    2.44 +  }
    2.45 +
    2.46    void movl(Register dest, Register src) {
    2.47      emit1(0x89, src.code, dest.code);
    2.48    }
    2.49 @@ -712,6 +743,74 @@
    2.50        outOfMemory_ = true;
    2.51    }
    2.52  
    2.53 +  void cpuid() {
    2.54 +    emit2(0x0f, 0xa2);
    2.55 +  }
    2.56 +
    2.57 +
    2.58 +  // SSE operations can only be used if the feature detection function has
    2.59 +  // been run *and* detected the appropriate level of functionality.
    2.60 +  void movss(FloatRegister dest, const Operand &src) {
    2.61 +    assert(Features().sse);
    2.62 +    emit3(0xf3, 0x0f, 0x10, dest.code, src);
    2.63 +  }
    2.64 +  void cvttss2si(Register dest, Register src) {
    2.65 +    assert(Features().sse);
    2.66 +    emit3(0xf3, 0x0f, 0x2c, dest.code, src.code);
    2.67 +  }
    2.68 +  void cvttss2si(Register dest, const Operand &src) {
    2.69 +    assert(Features().sse);
    2.70 +    emit3(0xf3, 0x0f, 0x2c, dest.code, src);
    2.71 +  }
    2.72 +  void cvtss2si(Register dest, Register src) {
    2.73 +    assert(Features().sse);
    2.74 +    emit3(0xf3, 0x0f, 0x2d, dest.code, src.code);
    2.75 +  }
    2.76 +  void cvtss2si(Register dest, const Operand &src) {
    2.77 +    assert(Features().sse);
    2.78 +    emit3(0xf3, 0x0f, 0x2d, dest.code, src);
    2.79 +  }
    2.80 +  void cvtsi2ss(FloatRegister dest, Register src) {
    2.81 +    assert(Features().sse);
    2.82 +    emit3(0xf3, 0x0f, 0x2a, dest.code, src.code);
    2.83 +  }
    2.84 +  void cvtsi2ss(FloatRegister dest, const Operand &src) {
    2.85 +    assert(Features().sse);
    2.86 +    emit3(0xf3, 0x0f, 0x2a, dest.code, src);
    2.87 +  }
    2.88 +  void addss(FloatRegister dest, const Operand &src) {
    2.89 +    assert(Features().sse);
    2.90 +    emit3(0xf3, 0x0f, 0x58, dest.code, src);
    2.91 +  }
    2.92 +  void subss(FloatRegister dest, const Operand &src) {
    2.93 +    assert(Features().sse);
    2.94 +    emit3(0xf3, 0x0f, 0x5c, dest.code, src);
    2.95 +  }
    2.96 +  void mulss(FloatRegister dest, const Operand &src) {
    2.97 +    assert(Features().sse);
    2.98 +    emit3(0xf3, 0x0f, 0x59, dest.code, src);
    2.99 +  }
   2.100 +  void divss(FloatRegister dest, const Operand &src) {
   2.101 +    assert(Features().sse);
   2.102 +    emit3(0xf3, 0x0f, 0x5e, dest.code, src);
   2.103 +  }
   2.104 +  void ucomiss(FloatRegister left, Register right) {
   2.105 +    emit2(0x0f, 0x2e, left.code, right.code);
   2.106 +  }
   2.107 +  void ucomiss(FloatRegister left, const Operand &right) {
   2.108 +    emit2(0x0f, 0x2e, left.code, right);
   2.109 +  }
   2.110 +
   2.111 +  // SSE2-only instructions.
   2.112 +  void movd(Register dest, FloatRegister src) {
   2.113 +    assert(Features().sse2);
   2.114 +    emit3(0x66, 0x0f, 0x7e, dest.code, src.code);
   2.115 +  }
   2.116 +  void movd(Register dest, const Operand &src) {
   2.117 +    assert(Features().sse2);
   2.118 +    emit3(0x66, 0x0f, 0x7e, dest.code, src);
   2.119 +  }
   2.120 +
   2.121    static void PatchRel32Absolute(uint8_t *ip, void *ptr) {
   2.122      int32_t delta = uint32_t(ptr) - uint32_t(ip);
   2.123      *reinterpret_cast<int32_t *>(ip - 4) = delta;
   2.124 @@ -806,6 +905,22 @@
   2.125      emit(reg, operand);
   2.126    }
   2.127  
   2.128 +  void emit3(uint8_t prefix1, uint8_t prefix2, uint8_t opcode) {
   2.129 +    ensureSpace();
   2.130 +    *pos_++ = prefix1;
   2.131 +    *pos_++ = prefix2;
   2.132 +    *pos_++ = opcode;
   2.133 +  }
   2.134 +  void emit3(uint8_t prefix1, uint8_t prefix2, uint8_t opcode, uint8_t reg, uint8_t opreg) {
   2.135 +    emit3(prefix1, prefix2, opcode);
   2.136 +    assert(reg <= 7);
   2.137 +    *pos_++ = (kModeReg << 6) | (reg << 3) | opreg;
   2.138 +  }
   2.139 +  void emit3(uint8_t prefix1, uint8_t prefix2, uint8_t opcode, uint8_t reg, const Operand &operand) {
   2.140 +    emit3(prefix1, prefix2, opcode);
   2.141 +    emit(reg, operand);
   2.142 +  }
   2.143 +
   2.144    template <typename T>
   2.145    void shift_cl(const T &t, uint8_t r) {
   2.146      emit1(0xd3, r, t);

     3.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     3.2 +++ b/public/jit/x86/macro-assembler-x86.h	Thu Aug 08 20:26:36 2013 -0700
     3.3 @@ -0,0 +1,104 @@
     3.4 +/**
     3.5 + * vim: set ts=8 sts=2 sw=2 tw=99 et:
     3.6 + * =============================================================================
     3.7 + * SourcePawn JIT SDK
     3.8 + * Copyright (C) 2004-2008 AlliedModders LLC.  All rights reserved.
     3.9 + * =============================================================================
    3.10 + *
    3.11 + * This program is free software; you can redistribute it and/or modify it under
    3.12 + * the terms of the GNU General Public License, version 3.0, as published by the
    3.13 + * Free Software Foundation.
    3.14 + * 
    3.15 + * This program is distributed in the hope that it will be useful, but WITHOUT
    3.16 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
    3.17 + * FOR A PARTICULAR PURPOSE.  See the GNU General Public License for more
    3.18 + * details.
    3.19 + *
    3.20 + * You should have received a copy of the GNU General Public License along with
    3.21 + * this program.  If not, see <http://www.gnu.org/licenses/>.
    3.22 + *
    3.23 + * As a special exception, AlliedModders LLC gives you permission to link the
    3.24 + * code of this program (as well as its derivative works) to "Half-Life 2," the
    3.25 + * "Source Engine," the "SourcePawn JIT," and any Game MODs that run on software
    3.26 + * by the Valve Corporation.  You must obey the GNU General Public License in
    3.27 + * all respects for all other code used.  Additionally, AlliedModders LLC grants
    3.28 + * this exception to all derivative works.  AlliedModders LLC defines further
    3.29 + * exceptions, found in LICENSE.txt (as of this writing, version JULY-31-2007),
    3.30 + * or <http://www.sourcemod.net/license.php>.
    3.31 + *
    3.32 + * Version: $Id$
    3.33 + */
    3.34 +#ifndef _include_sourcepawn_macroassembler_x86h__
    3.35 +#define _include_sourcepawn_macroassembler_x86h__
    3.36 +
    3.37 +#include <assembler.h>
    3.38 +#include <ke_vector.h>
    3.39 +#include <string.h>
    3.40 +#include <assembler-x86.h>
    3.41 +
    3.42 +class MacroAssemblerX86 : public AssemblerX86
    3.43 +{
    3.44 +  public:
    3.45 +    static void GenerateFeatureDetection(MacroAssemblerX86 &masm) {
    3.46 +      masm.push(ebp);
    3.47 +      masm.movl(ebp, esp);
    3.48 +      masm.push(ebx);
    3.49 +      {
    3.50 +        // Get ECX, EDX feature bits at the first CPUID level.
    3.51 +        masm.movl(eax, 1);
    3.52 +        masm.cpuid();
    3.53 +        masm.movl(eax, Operand(ebp, 8));
    3.54 +        masm.movl(Operand(eax, 0), ecx);
    3.55 +        masm.movl(eax, Operand(ebp, 12));
    3.56 +        masm.movl(Operand(eax, 0), edx);
    3.57 +      }
    3.58 +
    3.59 +      // Zero out bits we're not guaranteed to get.
    3.60 +      masm.movl(eax, Operand(ebp, 16));
    3.61 +      masm.movl(Operand(eax, 0), 0);
    3.62 +
    3.63 +      Label skip_level_7;
    3.64 +      {
    3.65 +        // Get EBX feature bits at 7th CPUID level.
    3.66 +        masm.movl(eax, 0);
    3.67 +        masm.cpuid();
    3.68 +        masm.cmpl(eax, 7);
    3.69 +        masm.j(below, &skip_level_7);
    3.70 +        masm.movl(eax, 7);
    3.71 +        masm.movl(ecx, 0);
    3.72 +        masm.cpuid();
    3.73 +        masm.movl(eax, Operand(ebp, 16));
    3.74 +        masm.movl(Operand(eax, 0), ebx);
    3.75 +      }
    3.76 +      masm.bind(&skip_level_7);
    3.77 +
    3.78 +      masm.pop(ebx);
    3.79 +      masm.pop(ebp);
    3.80 +      masm.ret();
    3.81 +    }
    3.82 +
    3.83 +    static void RunFeatureDetection(void *code) {
    3.84 +      typedef void (*fn_t)(int *reg_ecx, int *reg_edx, int *reg_ebx);
    3.85 +
    3.86 +      int reg_ecx, reg_edx, reg_ebx;
    3.87 +      ((fn_t)code)(&reg_ecx, &reg_edx, &reg_ebx);
    3.88 +      
    3.89 +      CPUFeatures features;
    3.90 +      features.fpu = !!(reg_edx & (1 << 0));
    3.91 +      features.mmx = !!(reg_edx & (1 << 23));
    3.92 +      features.sse = !!(reg_edx & (1 << 25));
    3.93 +      features.sse2 = !!(reg_edx & (1 << 26));
    3.94 +      features.sse3 = !!(reg_ecx & (1 << 0));
    3.95 +      features.ssse3 = !!(reg_ecx & (1 << 9));
    3.96 +      features.sse4_1 = !!(reg_ecx & (1 << 19));
    3.97 +      features.sse4_2 = !!(reg_ecx & (1 << 20));
    3.98 +      features.avx = !!(reg_ecx & (1 << 28));
    3.99 +      features.avx2 = !!(reg_ebx & (1 << 5));
   3.100 +      SetFeatures(features);
   3.101 +    }
   3.102 +
   3.103 +  private:
   3.104 +};
   3.105 +
   3.106 +#endif // _include_sourcepawn_macroassembler_x86h__
   3.107 +

     4.1 --- a/sourcepawn/jit/AMBuilder	Thu Aug 08 20:21:37 2013 -0700
     4.2 +++ b/sourcepawn/jit/AMBuilder	Thu Aug 08 20:26:36 2013 -0700
     4.3 @@ -37,7 +37,8 @@
     4.4  	'zlib/uncompr.c',
     4.5  	'zlib/zutil.c',
     4.6  	'md5/md5.cpp',
     4.7 -	'../../knight/shared/KeCodeAllocator.cpp'
     4.8 +	'../../knight/shared/KeCodeAllocator.cpp',
     4.9 +	'../../public/jit/x86/assembler-x86.cpp'
    4.10  	])
    4.11  SM.AutoVersion('sourcepawn/jit', binary)
    4.12  SM.ExtractDebugInfo(extension, binary)

     5.1 --- a/sourcepawn/jit/Makefile.shell	Thu Aug 08 20:21:37 2013 -0700
     5.2 +++ b/sourcepawn/jit/Makefile.shell	Thu Aug 08 20:26:36 2013 -0700
     5.3 @@ -34,6 +34,7 @@
     5.4  	zlib/zutil.c		\
     5.5  
     5.6  OBJECTS += ../../knight/shared/KeCodeAllocator.cpp
     5.7 +OBJECTS += ../../public/jit/x86/assembler-x86.cpp
     5.8  
     5.9  ##############################################
    5.10  ### CONFIGURE ANY OTHER FLAGS/OPTIONS HERE ###
    5.11 @@ -75,6 +76,7 @@
    5.12  endif
    5.13  
    5.14  OBJ_LINUX := $(OBJECTS:../../knight/shared/%.cpp=$(BIN_DIR)/knight/%.o)
    5.15 +OBJ_LINUX := $(OBJ_LINUX:../../public/jit/x86/%.cpp=$(BIN_DIR)/%.o)
    5.16  OBJ_LINUX := $(OBJ_LINUX:%.cpp=$(BIN_DIR)/%.o)
    5.17  OBJ_LINUX := $(OBJ_LINUX:%.c=$(BIN_DIR)/%.o)
    5.18  
    5.19 @@ -89,6 +91,9 @@
    5.20  $(BIN_DIR)/knight/%.o: ../../knight/shared/%.cpp
    5.21  	$(CXX) $(INCLUDE) $(CFLAGS) $(CXXFLAGS) -o $@ -c $<
    5.22  
    5.23 +$(BIN_DIR)/assembler-x86.o: ../../public/jit/x86/assembler-x86.cpp
    5.24 +	$(CXX) $(INCLUDE) $(CFLAGS) $(CXXFLAGS) -o $@ -c $<
    5.25 +
    5.26  all:
    5.27  	mkdir -p $(BIN_DIR)/x86
    5.28  	mkdir -p $(BIN_DIR)/md5

     6.1 --- a/sourcepawn/jit/x86/jit_x86.cpp	Thu Aug 08 20:21:37 2013 -0700
     6.2 +++ b/sourcepawn/jit/x86/jit_x86.cpp	Thu Aug 08 20:26:36 2013 -0700
     6.3 @@ -983,10 +983,15 @@
     6.4        break;
     6.5  
     6.6      case OP_FLOAT:
     6.7 -      __ fild32(Operand(edi, 0));
     6.8 -      __ subl(esp, 4);
     6.9 -      __ fstp32(Operand(esp, 0));
    6.10 -      __ pop(pri);
    6.11 +      if (MacroAssemblerX86::Features().sse2) {
    6.12 +        __ cvtsi2ss(xmm0, Operand(edi, 0));
    6.13 +        __ movd(pri, xmm0);
    6.14 +      } else {
    6.15 +        __ fild32(Operand(edi, 0));
    6.16 +        __ subl(esp, 4);
    6.17 +        __ fstp32(Operand(esp, 0));
    6.18 +        __ pop(pri);
    6.19 +      }
    6.20        __ addl(stk, 4);
    6.21        break;
    6.22  
    6.23 @@ -994,34 +999,52 @@
    6.24      case OP_FLOATSUB:
    6.25      case OP_FLOATMUL:
    6.26      case OP_FLOATDIV:
    6.27 -      __ subl(esp, 4);
    6.28 -      __ fld32(Operand(edi, 0));
    6.29 +      if (MacroAssemblerX86::Features().sse2) {
    6.30 +        __ movss(xmm0, Operand(stk, 0));
    6.31 +        if (op == OP_FLOATADD)
    6.32 +          __ addss(xmm0, Operand(stk, 4));
    6.33 +        else if (op == OP_FLOATSUB)
    6.34 +          __ subss(xmm0, Operand(stk, 4));
    6.35 +        else if (op == OP_FLOATMUL)
    6.36 +          __ mulss(xmm0, Operand(stk, 4));
    6.37 +        else if (op == OP_FLOATDIV)
    6.38 +          __ divss(xmm0, Operand(stk, 4));
    6.39 +        __ movd(pri, xmm0);
    6.40 +      } else {
    6.41 +        __ subl(esp, 4);
    6.42 +        __ fld32(Operand(stk, 0));
    6.43  
    6.44 -      if (op == OP_FLOATADD)
    6.45 -        __ fadd32(Operand(edi, 4));
    6.46 -      else if (op == OP_FLOATSUB)
    6.47 -        __ fsub32(Operand(edi, 4));
    6.48 -      else if (op == OP_FLOATMUL)
    6.49 -        __ fmul32(Operand(edi, 4));
    6.50 -      else if (op == OP_FLOATDIV)
    6.51 -        __ fdiv32(Operand(edi, 4));
    6.52 +        if (op == OP_FLOATADD)
    6.53 +          __ fadd32(Operand(stk, 4));
    6.54 +        else if (op == OP_FLOATSUB)
    6.55 +          __ fsub32(Operand(stk, 4));
    6.56 +        else if (op == OP_FLOATMUL)
    6.57 +          __ fmul32(Operand(stk, 4));
    6.58 +        else if (op == OP_FLOATDIV)
    6.59 +          __ fdiv32(Operand(stk, 4));
    6.60  
    6.61 -      __ fstp32(Operand(esp, 0));
    6.62 -      __ pop(pri);
    6.63 +        __ fstp32(Operand(esp, 0));
    6.64 +        __ pop(pri);
    6.65 +      }
    6.66        __ addl(stk, 8);
    6.67        break;
    6.68  
    6.69      case OP_RND_TO_NEAREST:
    6.70      {
    6.71 -      static float kRoundToNearest = 0.5f;
    6.72 -      // From http://wurstcaptures.untergrund.net/assembler_tricks.html#fastfloorf
    6.73 -      __ fld32(Operand(edi, 0));
    6.74 -      __ fadd32(st0, st0);
    6.75 -      __ fadd32(Operand(ExternalAddress(&kRoundToNearest)));
    6.76 -      __ subl(esp, 4);
    6.77 -      __ fistp32(Operand(esp, 0));
    6.78 -      __ pop(pri);
    6.79 -      __ sarl(pri, 1);
    6.80 +      if (MacroAssemblerX86::Features().sse) {
    6.81 +        // Assume no one is touching MXCSR.
    6.82 +        __ cvtss2si(pri, Operand(stk, 0));
    6.83 +      } else {
    6.84 +        static float kRoundToNearest = 0.5f;
    6.85 +        // From http://wurstcaptures.untergrund.net/assembler_tricks.html#fastfloorf
    6.86 +        __ fld32(Operand(stk, 0));
    6.87 +        __ fadd32(st0, st0);
    6.88 +        __ fadd32(Operand(ExternalAddress(&kRoundToNearest)));
    6.89 +        __ subl(esp, 4);
    6.90 +        __ fistp32(Operand(esp, 0));
    6.91 +        __ pop(pri);
    6.92 +        __ sarl(pri, 1);
    6.93 +      }
    6.94        __ addl(stk, 4);
    6.95        break;
    6.96      }
    6.97 @@ -1030,7 +1053,7 @@
    6.98      {
    6.99        static float kRoundToCeil = -0.5f;
   6.100        // From http://wurstcaptures.untergrund.net/assembler_tricks.html#fastfloorf
   6.101 -      __ fld32(Operand(edi, 0));
   6.102 +      __ fld32(Operand(stk, 0));
   6.103        __ fadd32(st0, st0);
   6.104        __ fsubr32(Operand(ExternalAddress(&kRoundToCeil)));
   6.105        __ subl(esp, 4);
   6.106 @@ -1043,15 +1066,19 @@
   6.107      }
   6.108  
   6.109      case OP_RND_TO_ZERO:
   6.110 -      __ fld32(Operand(edi, 0));
   6.111 -      __ subl(esp, 8);
   6.112 -      __ fstcw(Operand(esp, 4));
   6.113 -      __ movl(Operand(esp, 0), 0xfff);
   6.114 -      __ fldcw(Operand(esp, 0));
   6.115 -      __ fistp32(Operand(esp, 0));
   6.116 -      __ pop(pri);
   6.117 -      __ fldcw(Operand(esp, 0));
   6.118 -      __ addl(esp, 4);
   6.119 +      if (MacroAssemblerX86::Features().sse) {
   6.120 +        __ cvttss2si(pri, Operand(stk, 0));
   6.121 +      } else {
   6.122 +        __ fld32(Operand(stk, 0));
   6.123 +        __ subl(esp, 8);
   6.124 +        __ fstcw(Operand(esp, 4));
   6.125 +        __ movl(Operand(esp, 0), 0xfff);
   6.126 +        __ fldcw(Operand(esp, 0));
   6.127 +        __ fistp32(Operand(esp, 0));
   6.128 +        __ pop(pri);
   6.129 +        __ fldcw(Operand(esp, 0));
   6.130 +        __ addl(esp, 4);
   6.131 +      }
   6.132        __ addl(stk, 4);
   6.133        break;
   6.134  
   6.135 @@ -1071,10 +1098,15 @@
   6.136      case OP_FLOATCMP:
   6.137      {
   6.138        Label bl, ab, done;
   6.139 -      __ fld32(Operand(edi, 0));
   6.140 -      __ fld32(Operand(edi, 4));
   6.141 -      __ fucomip(st1);
   6.142 -      __ fstp(st0);
   6.143 +      if (MacroAssemblerX86::Features().sse) {
   6.144 +        __ movss(xmm0, Operand(stk, 4));
   6.145 +        __ ucomiss(xmm0, Operand(stk, 0));
   6.146 +      } else {
   6.147 +        __ fld32(Operand(stk, 0));
   6.148 +        __ fld32(Operand(stk, 4));
   6.149 +        __ fucomip(st1);
   6.150 +        __ fstp(st0);
   6.151 +      }
   6.152        __ j(above, &ab);
   6.153        __ j(below, &bl);
   6.154        __ xorl(pri, pri);
   6.155 @@ -1869,6 +1901,14 @@
   6.156    if (!m_pJitGenArray)
   6.157      return false;
   6.158  
   6.159 +  MacroAssemblerX86 masm;
   6.160 +  MacroAssemblerX86::GenerateFeatureDetection(masm);
   6.161 +  void *code = LinkCode(masm);
   6.162 +  if (!code)
   6.163 +    return false;
   6.164 +  MacroAssemblerX86::RunFeatureDetection(code);
   6.165 +  KE_FreeCode(g_pCodeCache, code);
   6.166 +
   6.167    return true;
   6.168  }
   6.169  

     7.1 --- a/sourcepawn/jit/x86/jit_x86.h	Thu Aug 08 20:21:37 2013 -0700
     7.2 +++ b/sourcepawn/jit/x86/jit_x86.h	Thu Aug 08 20:26:36 2013 -0700
     7.3 @@ -35,7 +35,7 @@
     7.4  #include <sp_vm_types.h>
     7.5  #include <sp_vm_api.h>
     7.6  #include <KeCodeAllocator.h>
     7.7 -#include <assembler-x86.h>
     7.8 +#include <macro-assembler-x86.h>
     7.9  #include <ke_vector.h>
    7.10  #include "jit_shared.h"
    7.11  #include "BaseRuntime.h"