Merge "ART: Optimizing compiler intrinsics"

3 files changed, 111 insertions, 6 deletions

diff --git a/compiler/utils/x86_64/assembler_x86_64.cc b/compiler/utils/x86_64/assembler_x86_64.cc
index d843a7213f..c7414a12fc 100644
--- a/compiler/utils/x86_64/assembler_x86_64.cc
+++ b/compiler/utils/x86_64/assembler_x86_64.cc
@@ -184,6 +184,20 @@ void X86_64Assembler::movl(const Address& dst, const Immediate& imm) {
   EmitImmediate(imm);
 }
 
+
+void X86_64Assembler::cmov(Condition c, CpuRegister dst, CpuRegister src) {
+  cmov(c, dst, src, true);
+}
+
+void X86_64Assembler::cmov(Condition c, CpuRegister dst, CpuRegister src, bool is64bit) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitOptionalRex(false, is64bit, dst.NeedsRex(), false, src.NeedsRex());
+  EmitUint8(0x0F);
+  EmitUint8(0x40 + c);
+  EmitRegisterOperand(dst.LowBits(), src.LowBits());
+}
+
+
 void X86_64Assembler::movzxb(CpuRegister dst, CpuRegister src) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitOptionalByteRegNormalizingRex32(dst, src);
@@ -369,19 +383,26 @@ void X86_64Assembler::movsxd(CpuRegister dst, const Address& src) {
 
 
 void X86_64Assembler::movd(XmmRegister dst, CpuRegister src) {
+  movd(dst, src, true);
+}
+
+void X86_64Assembler::movd(CpuRegister dst, XmmRegister src) {
+  movd(dst, src, true);
+}
+
+void X86_64Assembler::movd(XmmRegister dst, CpuRegister src, bool is64bit) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x66);
-  EmitRex64(dst, src);
+  EmitOptionalRex(false, is64bit, dst.NeedsRex(), false, src.NeedsRex());
   EmitUint8(0x0F);
   EmitUint8(0x6E);
   EmitOperand(dst.LowBits(), Operand(src));
 }
 
-
-void X86_64Assembler::movd(CpuRegister dst, XmmRegister src) {
+void X86_64Assembler::movd(CpuRegister dst, XmmRegister src, bool is64bit) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
   EmitUint8(0x66);
-  EmitRex64(src, dst);
+  EmitOptionalRex(false, is64bit, src.NeedsRex(), false, dst.NeedsRex());
   EmitUint8(0x0F);
   EmitUint8(0x7E);
   EmitOperand(src.LowBits(), Operand(dst));
@@ -826,6 +847,39 @@ void X86_64Assembler::andpd(XmmRegister dst, const Address& src) {
   EmitOperand(dst.LowBits(), src);
 }
 
+void X86_64Assembler::andpd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x54);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::andps(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x54);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::orpd(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitUint8(0x66);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x56);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
+
+void X86_64Assembler::orps(XmmRegister dst, XmmRegister src) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitOptionalRex32(dst, src);
+  EmitUint8(0x0F);
+  EmitUint8(0x56);
+  EmitXmmRegisterOperand(dst.LowBits(), src);
+}
 
 void X86_64Assembler::fldl(const Address& src) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
@@ -1757,6 +1811,20 @@ void X86_64Assembler::setcc(Condition condition, CpuRegister dst) {
   EmitUint8(0xC0 + dst.LowBits());
 }
 
+void X86_64Assembler::bswapl(CpuRegister dst) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitOptionalRex(false, false, false, false, dst.NeedsRex());
+  EmitUint8(0x0F);
+  EmitUint8(0xC8 + dst.LowBits());
+}
+
+void X86_64Assembler::bswapq(CpuRegister dst) {
+  AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+  EmitOptionalRex(false, true, false, false, dst.NeedsRex());
+  EmitUint8(0x0F);
+  EmitUint8(0xC8 + dst.LowBits());
+}
+
 
 void X86_64Assembler::LoadDoubleConstant(XmmRegister dst, double value) {
   // TODO: Need to have a code constants table.
diff --git a/compiler/utils/x86_64/assembler_x86_64.h b/compiler/utils/x86_64/assembler_x86_64.h
index ac8bc9ab49..5c8d608cf2 100644
--- a/compiler/utils/x86_64/assembler_x86_64.h
+++ b/compiler/utils/x86_64/assembler_x86_64.h
@@ -276,6 +276,9 @@ class X86_64Assembler FINAL : public Assembler {
   void movl(const Address& dst, CpuRegister src);
   void movl(const Address& dst, const Immediate& imm);
 
+  void cmov(Condition c, CpuRegister dst, CpuRegister src);  // This is the 64b version.
+  void cmov(Condition c, CpuRegister dst, CpuRegister src, bool is64bit);
+
   void movzxb(CpuRegister dst, CpuRegister src);
   void movzxb(CpuRegister dst, const Address& src);
   void movsxb(CpuRegister dst, CpuRegister src);
@@ -303,8 +306,10 @@ class X86_64Assembler FINAL : public Assembler {
   void movsxd(CpuRegister dst, CpuRegister src);
   void movsxd(CpuRegister dst, const Address& src);
 
-  void movd(XmmRegister dst, CpuRegister src);
-  void movd(CpuRegister dst, XmmRegister src);
+  void movd(XmmRegister dst, CpuRegister src);  // Note: this is the r64 version, formally movq.
+  void movd(CpuRegister dst, XmmRegister src);  // Note: this is the r64 version, formally movq.
+  void movd(XmmRegister dst, CpuRegister src, bool is64bit);
+  void movd(CpuRegister dst, XmmRegister src, bool is64bit);
 
   void addss(XmmRegister dst, XmmRegister src);
   void addss(XmmRegister dst, const Address& src);
@@ -360,6 +365,11 @@ class X86_64Assembler FINAL : public Assembler {
   void xorps(XmmRegister dst, XmmRegister src);
 
   void andpd(XmmRegister dst, const Address& src);
+  void andpd(XmmRegister dst, XmmRegister src);
+  void andps(XmmRegister dst, XmmRegister src);
+
+  void orpd(XmmRegister dst, XmmRegister src);
+  void orps(XmmRegister dst, XmmRegister src);
 
   void flds(const Address& src);
   void fstps(const Address& dst);
@@ -504,6 +514,9 @@ class X86_64Assembler FINAL : public Assembler {
 
   void setcc(Condition condition, CpuRegister dst);
 
+  void bswapl(CpuRegister dst);
+  void bswapq(CpuRegister dst);
+
   //
   // Macros for High-level operations.
   //
diff --git a/compiler/utils/x86_64/assembler_x86_64_test.cc b/compiler/utils/x86_64/assembler_x86_64_test.cc
index e93f45c0ef..6df4144004 100644
--- a/compiler/utils/x86_64/assembler_x86_64_test.cc
+++ b/compiler/utils/x86_64/assembler_x86_64_test.cc
@@ -689,6 +689,22 @@ TEST_F(AssemblerX86_64Test, Xorpd) {
   DriverStr(RepeatFF(&x86_64::X86_64Assembler::xorpd, "xorpd %{reg2}, %{reg1}"), "xorpd");
 }
 
+TEST_F(AssemblerX86_64Test, Andps) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::andps, "andps %{reg2}, %{reg1}"), "andps");
+}
+
+TEST_F(AssemblerX86_64Test, Andpd) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::andpd, "andpd %{reg2}, %{reg1}"), "andpd");
+}
+
+TEST_F(AssemblerX86_64Test, Orps) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::orps, "orps %{reg2}, %{reg1}"), "orps");
+}
+
+TEST_F(AssemblerX86_64Test, Orpd) {
+  DriverStr(RepeatFF(&x86_64::X86_64Assembler::orpd, "orpd %{reg2}, %{reg1}"), "orpd");
+}
+
 // X87
 
 std::string x87_fn(AssemblerX86_64Test::Base* assembler_test ATTRIBUTE_UNUSED,
@@ -757,6 +773,14 @@ TEST_F(AssemblerX86_64Test, RetAndLeave) {
 // MISC //
 //////////
 
+TEST_F(AssemblerX86_64Test, Bswapl) {
+  DriverStr(Repeatr(&x86_64::X86_64Assembler::bswapl, "bswap %{reg}"), "bswapl");
+}
+
+TEST_F(AssemblerX86_64Test, Bswapq) {
+  DriverStr(RepeatR(&x86_64::X86_64Assembler::bswapq, "bswap %{reg}"), "bswapq");
+}
+
 std::string setcc_test_fn(AssemblerX86_64Test::Base* assembler_test,
                           x86_64::X86_64Assembler* assembler) {
   // From Condition