7121648: Use 3-operands SIMD instructions on x86 with AVX

Summary: Use 3-operands SIMD instructions in C2 generated code for machines with AVX.
Reviewed-by: never

make/bsd/makefiles/adlc.make

@@ -39,9 +39,16 @@ OS = $(Platform_os_family)
 
 SOURCE.AD = $(OUTDIR)/$(OS)_$(Platform_arch_model).ad 
 
-SOURCES.AD = \
+ifeq ("${Platform_arch_model}", "${Platform_arch}")
+  SOURCES.AD = \
   $(call altsrc-replace,$(HS_COMMON_SRC)/cpu/$(ARCH)/vm/$(Platform_arch_model).ad) \
   $(call altsrc-replace,$(HS_COMMON_SRC)/os_cpu/$(OS)_$(ARCH)/vm/$(OS)_$(Platform_arch_model).ad)
+else
+  SOURCES.AD = \
+  $(call altsrc-replace,$(HS_COMMON_SRC)/cpu/$(ARCH)/vm/$(Platform_arch_model).ad) \
+  $(call altsrc-replace,$(HS_COMMON_SRC)/cpu/$(ARCH)/vm/$(Platform_arch).ad) \
+  $(call altsrc-replace,$(HS_COMMON_SRC)/os_cpu/$(OS)_$(ARCH)/vm/$(OS)_$(Platform_arch_model).ad)
+endif
 
 EXEC	= $(OUTDIR)/adlc
 

make/linux/makefiles/adlc.make

@@ -39,9 +39,16 @@ OS = $(Platform_os_family)
 
 SOURCE.AD = $(OUTDIR)/$(OS)_$(Platform_arch_model).ad 
 
-SOURCES.AD = \
+ifeq ("${Platform_arch_model}", "${Platform_arch}")
+  SOURCES.AD = \
   $(call altsrc-replace,$(HS_COMMON_SRC)/cpu/$(ARCH)/vm/$(Platform_arch_model).ad) \
   $(call altsrc-replace,$(HS_COMMON_SRC)/os_cpu/$(OS)_$(ARCH)/vm/$(OS)_$(Platform_arch_model).ad)
+else
+  SOURCES.AD = \
+  $(call altsrc-replace,$(HS_COMMON_SRC)/cpu/$(ARCH)/vm/$(Platform_arch_model).ad) \
+  $(call altsrc-replace,$(HS_COMMON_SRC)/cpu/$(ARCH)/vm/$(Platform_arch).ad) \
+  $(call altsrc-replace,$(HS_COMMON_SRC)/os_cpu/$(OS)_$(ARCH)/vm/$(OS)_$(Platform_arch_model).ad)
+endif
 
 EXEC	= $(OUTDIR)/adlc
 

make/solaris/makefiles/adlc.make

@@ -40,9 +40,16 @@ OS = $(Platform_os_family)
 
 SOURCE.AD = $(OUTDIR)/$(OS)_$(Platform_arch_model).ad 
 
-SOURCES.AD = \
+ifeq ("${Platform_arch_model}", "${Platform_arch}")
+  SOURCES.AD = \
   $(call altsrc-replace,$(HS_COMMON_SRC)/cpu/$(ARCH)/vm/$(Platform_arch_model).ad) \
   $(call altsrc-replace,$(HS_COMMON_SRC)/os_cpu/$(OS)_$(ARCH)/vm/$(OS)_$(Platform_arch_model).ad)
+else
+  SOURCES.AD = \
+  $(call altsrc-replace,$(HS_COMMON_SRC)/cpu/$(ARCH)/vm/$(Platform_arch_model).ad) \
+  $(call altsrc-replace,$(HS_COMMON_SRC)/cpu/$(ARCH)/vm/$(Platform_arch).ad) \
+  $(call altsrc-replace,$(HS_COMMON_SRC)/os_cpu/$(OS)_$(ARCH)/vm/$(OS)_$(Platform_arch_model).ad)
+endif
 
 EXEC	= $(OUTDIR)/adlc
 

make/windows/makefiles/adlc.make

@@ -53,6 +53,17 @@ CPP_INCLUDE_DIRS=\
   /I "$(WorkSpace)\src\os\windows\vm" \
   /I "$(WorkSpace)\src\cpu\$(Platform_arch)\vm"
 
+!if "$(Platform_arch_model)" == "$(Platform_arch)"
+SOURCES_AD=\
+  $(WorkSpace)/src/cpu/$(Platform_arch)/vm/$(Platform_arch_model).ad \
+  $(WorkSpace)/src/os_cpu/windows_$(Platform_arch)/vm/windows_$(Platform_arch_model).ad
+!else
+SOURCES_AD=\
+  $(WorkSpace)/src/cpu/$(Platform_arch)/vm/$(Platform_arch_model).ad \
+  $(WorkSpace)/src/cpu/$(Platform_arch)/vm/$(Platform_arch).ad \
+  $(WorkSpace)/src/os_cpu/windows_$(Platform_arch)/vm/windows_$(Platform_arch_model).ad
+!endif
+
 # NOTE! If you add any files here, you must also update GENERATED_NAMES_IN_DIR
 # and ProjectCreatorIDEOptions in projectcreator.make. 
 GENERATED_NAMES=\
@@ -105,7 +116,6 @@ $(GENERATED_NAMES_IN_DIR): $(Platform_arch_model).ad adlc.exe
 	$(ADLC) $(ADLCFLAGS) $(Platform_arch_model).ad
 	mv $(GENERATED_NAMES) $(AdlcOutDir)/
 
-$(Platform_arch_model).ad: $(WorkSpace)/src/cpu/$(Platform_arch)/vm/$(Platform_arch_model).ad $(WorkSpace)/src/os_cpu/windows_$(Platform_arch)/vm/windows_$(Platform_arch_model).ad
+$(Platform_arch_model).ad: $(SOURCES_AD)
 	rm -f $(Platform_arch_model).ad
-	cat $(WorkSpace)/src/cpu/$(Platform_arch)/vm/$(Platform_arch_model).ad  \
-	    $(WorkSpace)/src/os_cpu/windows_$(Platform_arch)/vm/windows_$(Platform_arch_model).ad >$(Platform_arch_model).ad
+	cat $(SOURCES_AD) >$(Platform_arch_model).ad

src/cpu/x86/vm/assembler_x86.cpp

@@ -2932,6 +2932,161 @@ void Assembler::xorps(XMMRegister dst, Address src) {
   emit_operand(dst, src);
 }
 
+// AVX 3-operands non destructive source instructions (encoded with VEX prefix)
+
+void Assembler::vaddsd(XMMRegister dst, XMMRegister nds, Address src) {
+  assert(VM_Version::supports_avx(), "");
+  InstructionMark im(this);
+  vex_prefix(dst, nds, src, VEX_SIMD_F2);
+  emit_byte(0x58);
+  emit_operand(dst, src);
+}
+
+void Assembler::vaddsd(XMMRegister dst, XMMRegister nds, XMMRegister src) {
+  assert(VM_Version::supports_avx(), "");
+  int encode = vex_prefix_and_encode(dst, nds, src, VEX_SIMD_F2);
+  emit_byte(0x58);
+  emit_byte(0xC0 | encode);
+}
+
+void Assembler::vaddss(XMMRegister dst, XMMRegister nds, Address src) {
+  assert(VM_Version::supports_avx(), "");
+  InstructionMark im(this);
+  vex_prefix(dst, nds, src, VEX_SIMD_F3);
+  emit_byte(0x58);
+  emit_operand(dst, src);
+}
+
+void Assembler::vaddss(XMMRegister dst, XMMRegister nds, XMMRegister src) {
+  assert(VM_Version::supports_avx(), "");
+  int encode = vex_prefix_and_encode(dst, nds, src, VEX_SIMD_F3);
+  emit_byte(0x58);
+  emit_byte(0xC0 | encode);
+}
+
+void Assembler::vandpd(XMMRegister dst, XMMRegister nds, Address src) {
+  assert(VM_Version::supports_avx(), "");
+  InstructionMark im(this);
+  vex_prefix(dst, nds, src, VEX_SIMD_66); // 128-bit vector
+  emit_byte(0x54);
+  emit_operand(dst, src);
+}
+
+void Assembler::vandps(XMMRegister dst, XMMRegister nds, Address src) {
+  assert(VM_Version::supports_avx(), "");
+  InstructionMark im(this);
+  vex_prefix(dst, nds, src, VEX_SIMD_NONE); // 128-bit vector
+  emit_byte(0x54);
+  emit_operand(dst, src);
+}
+
+void Assembler::vdivsd(XMMRegister dst, XMMRegister nds, Address src) {
+  assert(VM_Version::supports_avx(), "");
+  InstructionMark im(this);
+  vex_prefix(dst, nds, src, VEX_SIMD_F2);
+  emit_byte(0x5E);
+  emit_operand(dst, src);
+}
+
+void Assembler::vdivsd(XMMRegister dst, XMMRegister nds, XMMRegister src) {
+  assert(VM_Version::supports_avx(), "");
+  int encode = vex_prefix_and_encode(dst, nds, src, VEX_SIMD_F2);
+  emit_byte(0x5E);
+  emit_byte(0xC0 | encode);
+}
+
+void Assembler::vdivss(XMMRegister dst, XMMRegister nds, Address src) {
+  assert(VM_Version::supports_avx(), "");
+  InstructionMark im(this);
+  vex_prefix(dst, nds, src, VEX_SIMD_F3);
+  emit_byte(0x5E);
+  emit_operand(dst, src);
+}
+
+void Assembler::vdivss(XMMRegister dst, XMMRegister nds, XMMRegister src) {
+  assert(VM_Version::supports_avx(), "");
+  int encode = vex_prefix_and_encode(dst, nds, src, VEX_SIMD_F3);
+  emit_byte(0x5E);
+  emit_byte(0xC0 | encode);
+}
+
+void Assembler::vmulsd(XMMRegister dst, XMMRegister nds, Address src) {
+  assert(VM_Version::supports_avx(), "");
+  InstructionMark im(this);
+  vex_prefix(dst, nds, src, VEX_SIMD_F2);
+  emit_byte(0x59);
+  emit_operand(dst, src);
+}
+
+void Assembler::vmulsd(XMMRegister dst, XMMRegister nds, XMMRegister src) {
+  assert(VM_Version::supports_avx(), "");
+  int encode = vex_prefix_and_encode(dst, nds, src, VEX_SIMD_F2);
+  emit_byte(0x59);
+  emit_byte(0xC0 | encode);
+}
+
+void Assembler::vmulss(XMMRegister dst, XMMRegister nds, Address src) {
+  InstructionMark im(this);
+  vex_prefix(dst, nds, src, VEX_SIMD_F3);
+  emit_byte(0x59);
+  emit_operand(dst, src);
+}
+
+void Assembler::vmulss(XMMRegister dst, XMMRegister nds, XMMRegister src) {
+  assert(VM_Version::supports_avx(), "");
+  int encode = vex_prefix_and_encode(dst, nds, src, VEX_SIMD_F3);
+  emit_byte(0x59);
+  emit_byte(0xC0 | encode);
+}
+
+
+void Assembler::vsubsd(XMMRegister dst, XMMRegister nds, Address src) {
+  assert(VM_Version::supports_avx(), "");
+  InstructionMark im(this);
+  vex_prefix(dst, nds, src, VEX_SIMD_F2);
+  emit_byte(0x5C);
+  emit_operand(dst, src);
+}
+
+void Assembler::vsubsd(XMMRegister dst, XMMRegister nds, XMMRegister src) {
+  assert(VM_Version::supports_avx(), "");
+  int encode = vex_prefix_and_encode(dst, nds, src, VEX_SIMD_F2);
+  emit_byte(0x5C);
+  emit_byte(0xC0 | encode);
+}
+
+void Assembler::vsubss(XMMRegister dst, XMMRegister nds, Address src) {
+  assert(VM_Version::supports_avx(), "");
+  InstructionMark im(this);
+  vex_prefix(dst, nds, src, VEX_SIMD_F3);
+  emit_byte(0x5C);
+  emit_operand(dst, src);
+}
+
+void Assembler::vsubss(XMMRegister dst, XMMRegister nds, XMMRegister src) {
+  assert(VM_Version::supports_avx(), "");
+  int encode = vex_prefix_and_encode(dst, nds, src, VEX_SIMD_F3);
+  emit_byte(0x5C);
+  emit_byte(0xC0 | encode);
+}
+
+void Assembler::vxorpd(XMMRegister dst, XMMRegister nds, Address src) {
+  assert(VM_Version::supports_avx(), "");
+  InstructionMark im(this);
+  vex_prefix(dst, nds, src, VEX_SIMD_66); // 128-bit vector
+  emit_byte(0x57);
+  emit_operand(dst, src);
+}
+
+void Assembler::vxorps(XMMRegister dst, XMMRegister nds, Address src) {
+  assert(VM_Version::supports_avx(), "");
+  InstructionMark im(this);
+  vex_prefix(dst, nds, src, VEX_SIMD_NONE); // 128-bit vector
+  emit_byte(0x57);
+  emit_operand(dst, src);
+}
+
+
 #ifndef _LP64
 // 32bit only pieces of the assembler
 
@@ -7235,6 +7390,157 @@ void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) {
   }
 }
 
+void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) {
+  if (reachable(src)) {
+    Assembler::ucomisd(dst, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    Assembler::ucomisd(dst, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) {
+  if (reachable(src)) {
+    Assembler::ucomiss(dst, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    Assembler::ucomiss(dst, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) {
+  // Used in sign-bit flipping with aligned address.
+  assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
+  if (reachable(src)) {
+    Assembler::xorpd(dst, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    Assembler::xorpd(dst, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) {
+  // Used in sign-bit flipping with aligned address.
+  assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
+  if (reachable(src)) {
+    Assembler::xorps(dst, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    Assembler::xorps(dst, Address(rscratch1, 0));
+  }
+}
+
+// AVX 3-operands instructions
+
+void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
+  if (reachable(src)) {
+    vaddsd(dst, nds, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    vaddsd(dst, nds, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
+  if (reachable(src)) {
+    vaddss(dst, nds, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    vaddss(dst, nds, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
+  if (reachable(src)) {
+    vandpd(dst, nds, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    vandpd(dst, nds, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
+  if (reachable(src)) {
+    vandps(dst, nds, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    vandps(dst, nds, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
+  if (reachable(src)) {
+    vdivsd(dst, nds, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    vdivsd(dst, nds, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
+  if (reachable(src)) {
+    vdivss(dst, nds, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    vdivss(dst, nds, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
+  if (reachable(src)) {
+    vmulsd(dst, nds, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    vmulsd(dst, nds, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
+  if (reachable(src)) {
+    vmulss(dst, nds, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    vmulss(dst, nds, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
+  if (reachable(src)) {
+    vsubsd(dst, nds, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    vsubsd(dst, nds, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
+  if (reachable(src)) {
+    vsubss(dst, nds, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    vsubss(dst, nds, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
+  if (reachable(src)) {
+    vxorpd(dst, nds, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    vxorpd(dst, nds, Address(rscratch1, 0));
+  }
+}
+
+void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
+  if (reachable(src)) {
+    vxorps(dst, nds, as_Address(src));
+  } else {
+    lea(rscratch1, src);
+    vxorps(dst, nds, Address(rscratch1, 0));
+  }
+}
+
+
 //////////////////////////////////////////////////////////////////////////////////
 #ifndef SERIALGC
 
@@ -8119,46 +8425,6 @@ void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
 }
 
 
-void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) {
-  if (reachable(src)) {
-    Assembler::ucomisd(dst, as_Address(src));
-  } else {
-    lea(rscratch1, src);
-    Assembler::ucomisd(dst, Address(rscratch1, 0));
-  }
-}
-
-void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) {
-  if (reachable(src)) {
-    Assembler::ucomiss(dst, as_Address(src));
-  } else {
-    lea(rscratch1, src);
-    Assembler::ucomiss(dst, Address(rscratch1, 0));
-  }
-}
-
-void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) {
-  // Used in sign-bit flipping with aligned address.
-  assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
-  if (reachable(src)) {
-    Assembler::xorpd(dst, as_Address(src));
-  } else {
-    lea(rscratch1, src);
-    Assembler::xorpd(dst, Address(rscratch1, 0));
-  }
-}
-
-void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) {
-  // Used in sign-bit flipping with aligned address.
-  assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
-  if (reachable(src)) {
-    Assembler::xorps(dst, as_Address(src));
-  } else {
-    lea(rscratch1, src);
-    Assembler::xorps(dst, Address(rscratch1, 0));
-  }
-}
-
 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) {
   if (VM_Version::supports_cmov()) {
     cmovl(cc, dst, src);

src/cpu/x86/vm/assembler_x86.hpp

@@ -589,10 +589,21 @@ private:
                   VexSimdPrefix pre, VexOpcode opc,
                   bool vex_w, bool vector256);
 
+  void vex_prefix(XMMRegister dst, XMMRegister nds, Address src,
+                  VexSimdPrefix pre, bool vector256 = false) {
+     vex_prefix(src, nds->encoding(), dst->encoding(),
+                pre, VEX_OPCODE_0F, false, vector256);
+  }
+
   int  vex_prefix_and_encode(int dst_enc, int nds_enc, int src_enc,
                              VexSimdPrefix pre, VexOpcode opc,
                              bool vex_w, bool vector256);
 
+  int  vex_prefix_and_encode(XMMRegister dst, XMMRegister nds, XMMRegister src,
+                             VexSimdPrefix pre, bool vector256 = false) {
+     return vex_prefix_and_encode(dst->encoding(), nds->encoding(), src->encoding(),
+                                  pre, VEX_OPCODE_0F, false, vector256);
+  }
 
   void simd_prefix(XMMRegister xreg, XMMRegister nds, Address adr,
                    VexSimdPrefix pre, VexOpcode opc = VEX_OPCODE_0F,
@@ -1574,6 +1585,29 @@ private:
 
   void set_byte_if_not_zero(Register dst); // sets reg to 1 if not zero, otherwise 0
 
+  // AVX 3-operands instructions (encoded with VEX prefix)
+  void vaddsd(XMMRegister dst, XMMRegister nds, Address src);
+  void vaddsd(XMMRegister dst, XMMRegister nds, XMMRegister src);
+  void vaddss(XMMRegister dst, XMMRegister nds, Address src);
+  void vaddss(XMMRegister dst, XMMRegister nds, XMMRegister src);
+  void vandpd(XMMRegister dst, XMMRegister nds, Address src);
+  void vandps(XMMRegister dst, XMMRegister nds, Address src);
+  void vdivsd(XMMRegister dst, XMMRegister nds, Address src);
+  void vdivsd(XMMRegister dst, XMMRegister nds, XMMRegister src);
+  void vdivss(XMMRegister dst, XMMRegister nds, Address src);
+  void vdivss(XMMRegister dst, XMMRegister nds, XMMRegister src);
+  void vmulsd(XMMRegister dst, XMMRegister nds, Address src);
+  void vmulsd(XMMRegister dst, XMMRegister nds, XMMRegister src);
+  void vmulss(XMMRegister dst, XMMRegister nds, Address src);
+  void vmulss(XMMRegister dst, XMMRegister nds, XMMRegister src);
+  void vsubsd(XMMRegister dst, XMMRegister nds, Address src);
+  void vsubsd(XMMRegister dst, XMMRegister nds, XMMRegister src);
+  void vsubss(XMMRegister dst, XMMRegister nds, Address src);
+  void vsubss(XMMRegister dst, XMMRegister nds, XMMRegister src);
+  void vxorpd(XMMRegister dst, XMMRegister nds, Address src);
+  void vxorps(XMMRegister dst, XMMRegister nds, Address src);
+
+
  protected:
   // Next instructions require address alignment 16 bytes SSE mode.
   // They should be called only from corresponding MacroAssembler instructions.
@@ -2422,6 +2456,53 @@ public:
   void xorps(XMMRegister dst, Address src)     { Assembler::xorps(dst, src); }
   void xorps(XMMRegister dst, AddressLiteral src);
 
+  // AVX 3-operands instructions
+
+  void vaddsd(XMMRegister dst, XMMRegister nds, XMMRegister src) { Assembler::vaddsd(dst, nds, src); }
+  void vaddsd(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vaddsd(dst, nds, src); }
+  void vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+
+  void vaddss(XMMRegister dst, XMMRegister nds, XMMRegister src) { Assembler::vaddss(dst, nds, src); }
+  void vaddss(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vaddss(dst, nds, src); }
+  void vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+
+  void vandpd(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vandpd(dst, nds, src); }
+  void vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+
+  void vandps(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vandps(dst, nds, src); }
+  void vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+
+  void vdivsd(XMMRegister dst, XMMRegister nds, XMMRegister src) { Assembler::vdivsd(dst, nds, src); }
+  void vdivsd(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vdivsd(dst, nds, src); }
+  void vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+
+  void vdivss(XMMRegister dst, XMMRegister nds, XMMRegister src) { Assembler::vdivss(dst, nds, src); }
+  void vdivss(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vdivss(dst, nds, src); }
+  void vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+
+  void vmulsd(XMMRegister dst, XMMRegister nds, XMMRegister src) { Assembler::vmulsd(dst, nds, src); }
+  void vmulsd(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vmulsd(dst, nds, src); }
+  void vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+
+  void vmulss(XMMRegister dst, XMMRegister nds, XMMRegister src) { Assembler::vmulss(dst, nds, src); }
+  void vmulss(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vmulss(dst, nds, src); }
+  void vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+
+  void vsubsd(XMMRegister dst, XMMRegister nds, XMMRegister src) { Assembler::vsubsd(dst, nds, src); }
+  void vsubsd(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vsubsd(dst, nds, src); }
+  void vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+
+  void vsubss(XMMRegister dst, XMMRegister nds, XMMRegister src) { Assembler::vsubss(dst, nds, src); }
+  void vsubss(XMMRegister dst, XMMRegister nds, Address src)     { Assembler::vsubss(dst, nds, src); }
+  void vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+
+  void vxorpd(XMMRegister dst, XMMRegister nds, Address src) { Assembler::vxorpd(dst, nds, src); }
+  void vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+
+  void vxorps(XMMRegister dst, XMMRegister nds, Address src) { Assembler::vxorps(dst, nds, src); }
+  void vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src);
+
+
   // Data
 
   void cmov32( Condition cc, Register dst, Address  src);

src/share/vm/opto/matcher.cpp

@@ -1365,31 +1365,36 @@ static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool s
 
   const Type *t = m->bottom_type();
 
-  if( t->singleton() ) {
+  if (t->singleton()) {
     // Never force constants into registers.  Allow them to match as
     // constants or registers.  Copies of the same value will share
     // the same register.  See find_shared_node.
     return false;
   } else {                      // Not a constant
     // Stop recursion if they have different Controls.
-    // Slot 0 of constants is not really a Control.
-    if( control && m->in(0) && control != m->in(0) ) {
+    Node* m_control = m->in(0);
+    // Control of load's memory can post-dominates load's control.
+    // So use it since load can't float above its memory.
+    Node* mem_control = (m->is_Load()) ? m->in(MemNode::Memory)->in(0) : NULL;
+    if (control && m_control && control != m_control && control != mem_control) {
 
       // Actually, we can live with the most conservative control we
       // find, if it post-dominates the others.  This allows us to
       // pick up load/op/store trees where the load can float a little
       // above the store.
       Node *x = control;
-      const uint max_scan = 6;   // Arbitrary scan cutoff
+      const uint max_scan = 6;  // Arbitrary scan cutoff
       uint j;
-      for( j=0; j<max_scan; j++ ) {
-        if( x->is_Region() )    // Bail out at merge points
+      for (j=0; j<max_scan; j++) {
+        if (x->is_Region())     // Bail out at merge points
           return true;
         x = x->in(0);
-        if( x == m->in(0) )     // Does 'control' post-dominate
+        if (x == m_control)     // Does 'control' post-dominate
           break;                // m->in(0)?  If so, we can use it
+        if (x == mem_control)   // Does 'control' post-dominate
+          break;                // mem_control?  If so, we can use it
       }
-      if( j == max_scan )       // No post-domination before scan end?
+      if (j == max_scan)        // No post-domination before scan end?
         return true;            // Then break the match tree up
     }
     if (m->is_DecodeN() && Matcher::narrow_oop_use_complex_address()) {