diff --git a/compiler/utils/x86/assembler_x86.cc b/compiler/utils/x86/assembler_x86.cc
index bef32f8..b50f1af 100644
--- a/compiler/utils/x86/assembler_x86.cc
+++ b/compiler/utils/x86/assembler_x86.cc
@@ -1238,6 +1238,101 @@
   EmitXmmRegisterOperand(dst, src);
 }
 
+
+void X86Assembler::psadbw(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitUint8(0x0F);
+  EmitUint8(0xF6);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::pmaddwd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitUint8(0x0F);
+  EmitUint8(0xF5);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::phaddw(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitUint8(0x0F);
+  EmitUint8(0x38);
+  EmitUint8(0x01);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::phaddd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitUint8(0x0F);
+  EmitUint8(0x38);
+  EmitUint8(0x02);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::haddps(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0xF2);
+  EmitUint8(0x0F);
+  EmitUint8(0x7C);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::haddpd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitUint8(0x0F);
+  EmitUint8(0x7C);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::phsubw(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitUint8(0x0F);
+  EmitUint8(0x38);
+  EmitUint8(0x05);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::phsubd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitUint8(0x0F);
+  EmitUint8(0x38);
+  EmitUint8(0x06);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::hsubps(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0xF2);
+  EmitUint8(0x0F);
+  EmitUint8(0x7D);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::hsubpd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitUint8(0x0F);
+  EmitUint8(0x7D);
+  EmitXmmRegisterOperand(dst, src);
+}
+
+
 void X86Assembler::pminsb(XmmRegister dst, XmmRegister src) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x66);
diff --git a/compiler/utils/x86/assembler_x86.h b/compiler/utils/x86/assembler_x86.h
index c4bb9ee..8578340 100644
--- a/compiler/utils/x86/assembler_x86.h
+++ b/compiler/utils/x86/assembler_x86.h
@@ -497,6 +497,16 @@
 
   void pavgb(XmmRegister dst, XmmRegister src);  // no addr variant (for now)
   void pavgw(XmmRegister dst, XmmRegister src);
+  void psadbw(XmmRegister dst, XmmRegister src);
+  void pmaddwd(XmmRegister dst, XmmRegister src);
+  void phaddw(XmmRegister dst, XmmRegister src);
+  void phaddd(XmmRegister dst, XmmRegister src);
+  void haddps(XmmRegister dst, XmmRegister src);
+  void haddpd(XmmRegister dst, XmmRegister src);
+  void phsubw(XmmRegister dst, XmmRegister src);
+  void phsubd(XmmRegister dst, XmmRegister src);
+  void hsubps(XmmRegister dst, XmmRegister src);
+  void hsubpd(XmmRegister dst, XmmRegister src);
 
   void pminsb(XmmRegister dst, XmmRegister src);  // no addr variant (for now)
   void pmaxsb(XmmRegister dst, XmmRegister src);
diff --git a/compiler/utils/x86/assembler_x86_test.cc b/compiler/utils/x86/assembler_x86_test.cc
index 34f2a47..3e1244e 100644
--- a/compiler/utils/x86/assembler_x86_test.cc
+++ b/compiler/utils/x86/assembler_x86_test.cc
@@ -613,6 +613,46 @@
   DriverStr(RepeatFF(&x86::X86Assembler::pavgw, "pavgw %{reg2}, %{reg1}"), "pavgw");
 }
 
+TEST_F(AssemblerX86Test, PSadBW) {
+  DriverStr(RepeatFF(&x86::X86Assembler::psadbw, "psadbw %{reg2}, %{reg1}"), "psadbw");
+}
+
+TEST_F(AssemblerX86Test, PMAddWD) {
+  DriverStr(RepeatFF(&x86::X86Assembler::pmaddwd, "pmaddwd %{reg2}, %{reg1}"), "pmaddwd");
+}
+
+TEST_F(AssemblerX86Test, PHAddW) {
+  DriverStr(RepeatFF(&x86::X86Assembler::phaddw, "phaddw %{reg2}, %{reg1}"), "phaddw");
+}
+
+TEST_F(AssemblerX86Test, PHAddD) {
+  DriverStr(RepeatFF(&x86::X86Assembler::phaddd, "phaddd %{reg2}, %{reg1}"), "phaddd");
+}
+
+TEST_F(AssemblerX86Test, HAddPS) {
+  DriverStr(RepeatFF(&x86::X86Assembler::haddps, "haddps %{reg2}, %{reg1}"), "haddps");
+}
+
+TEST_F(AssemblerX86Test, HAddPD) {
+  DriverStr(RepeatFF(&x86::X86Assembler::haddpd, "haddpd %{reg2}, %{reg1}"), "haddpd");
+}
+
+TEST_F(AssemblerX86Test, PHSubW) {
+  DriverStr(RepeatFF(&x86::X86Assembler::phsubw, "phsubw %{reg2}, %{reg1}"), "phsubw");
+}
+
+TEST_F(AssemblerX86Test, PHSubD) {
+  DriverStr(RepeatFF(&x86::X86Assembler::phsubd, "phsubd %{reg2}, %{reg1}"), "phsubd");
+}
+
+TEST_F(AssemblerX86Test, HSubPS) {
+  DriverStr(RepeatFF(&x86::X86Assembler::hsubps, "hsubps %{reg2}, %{reg1}"), "hsubps");
+}
+
+TEST_F(AssemblerX86Test, HSubPD) {
+  DriverStr(RepeatFF(&x86::X86Assembler::hsubpd, "hsubpd %{reg2}, %{reg1}"), "hsubpd");
+}
+
 TEST_F(AssemblerX86Test, PMinSB) {
   DriverStr(RepeatFF(&x86::X86Assembler::pminsb, "pminsb %{reg2}, %{reg1}"), "pminsb");
 }
diff --git a/compiler/utils/x86_64/assembler_x86_64.cc b/compiler/utils/x86_64/assembler_x86_64.cc
index 82d1174..ea69a1c 100644
--- a/compiler/utils/x86_64/assembler_x86_64.cc
+++ b/compiler/utils/x86_64/assembler_x86_64.cc
@@ -1445,6 +1445,100 @@
   EmitXmmRegisterOperand(dst.LowBits(), src);
 }
 
+void X86_64Assembler::psadbw(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0xF6);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::pmaddwd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0xF5);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::phaddw(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x38);
+  EmitUint8(0x01);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::phaddd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x38);
+  EmitUint8(0x02);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::haddps(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0xF2);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x7C);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::haddpd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x7C);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::phsubw(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x38);
+  EmitUint8(0x05);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::phsubd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x38);
+  EmitUint8(0x06);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::hsubps(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0xF2);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x7D);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::hsubpd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x7D);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
 void X86_64Assembler::pminsb(XmmRegister dst, XmmRegister src) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x66);
diff --git a/compiler/utils/x86_64/assembler_x86_64.h b/compiler/utils/x86_64/assembler_x86_64.h
index 6e584fe..41450bf 100644
--- a/compiler/utils/x86_64/assembler_x86_64.h
+++ b/compiler/utils/x86_64/assembler_x86_64.h
@@ -525,6 +525,16 @@
 
   void pavgb(XmmRegister dst, XmmRegister src);  // no addr variant (for now)
   void pavgw(XmmRegister dst, XmmRegister src);
+  void psadbw(XmmRegister dst, XmmRegister src);
+  void pmaddwd(XmmRegister dst, XmmRegister src);
+  void phaddw(XmmRegister dst, XmmRegister src);
+  void phaddd(XmmRegister dst, XmmRegister src);
+  void haddps(XmmRegister dst, XmmRegister src);
+  void haddpd(XmmRegister dst, XmmRegister src);
+  void phsubw(XmmRegister dst, XmmRegister src);
+  void phsubd(XmmRegister dst, XmmRegister src);
+  void hsubps(XmmRegister dst, XmmRegister src);
+  void hsubpd(XmmRegister dst, XmmRegister src);
 
   void pminsb(XmmRegister dst, XmmRegister src);  // no addr variant (for now)
   void pmaxsb(XmmRegister dst, XmmRegister src);
diff --git a/compiler/utils/x86_64/assembler_x86_64_test.cc b/compiler/utils/x86_64/assembler_x86_64_test.cc
index b574003..ec14e7a 100644
--- a/compiler/utils/x86_64/assembler_x86_64_test.cc
+++ b/compiler/utils/x86_64/assembler_x86_64_test.cc
@@ -1301,6 +1301,46 @@
   DriverStr(RepeatFF(&x86_64::X86_64Assembler::pavgw, "pavgw %{reg2}, %{reg1}"), "pavgw");
 }
 
+TEST_F(AssemblerX86_64Test, Psadbw) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::psadbw, "psadbw %{reg2}, %{reg1}"), "psadbw");
+}
+
+TEST_F(AssemblerX86_64Test, Pmaddwd) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::pmaddwd, "pmaddwd %{reg2}, %{reg1}"), "pmadwd");
+}
+
+TEST_F(AssemblerX86_64Test, Phaddw) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::phaddw, "phaddw %{reg2}, %{reg1}"), "phaddw");
+}
+
+TEST_F(AssemblerX86_64Test, Phaddd) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::phaddd, "phaddd %{reg2}, %{reg1}"), "phaddd");
+}
+
+TEST_F(AssemblerX86_64Test, Haddps) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::haddps, "haddps %{reg2}, %{reg1}"), "haddps");
+}
+
+TEST_F(AssemblerX86_64Test, Haddpd) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::haddpd, "haddpd %{reg2}, %{reg1}"), "haddpd");
+}
+
+TEST_F(AssemblerX86_64Test, Phsubw) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::phsubw, "phsubw %{reg2}, %{reg1}"), "phsubw");
+}
+
+TEST_F(AssemblerX86_64Test, Phsubd) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::phsubd, "phsubd %{reg2}, %{reg1}"), "phsubd");
+}
+
+TEST_F(AssemblerX86_64Test, Hsubps) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::hsubps, "hsubps %{reg2}, %{reg1}"), "hsubps");
+}
+
+TEST_F(AssemblerX86_64Test, Hsubpd) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::hsubpd, "hsubpd %{reg2}, %{reg1}"), "hsubpd");
+}
+
 TEST_F(AssemblerX86_64Test, Pminsb) {
   DriverStr(RepeatFF(&x86_64::X86_64Assembler::pminsb, "pminsb %{reg2}, %{reg1}"), "pminsb");
 }