Merge branch 'dev'v7.03

12 files changed, 259 insertions, 123 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index e61de1f..182a5de 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.5)
 
-project(xbyak LANGUAGES CXX VERSION 7.02)
+project(xbyak LANGUAGES CXX VERSION 7.03)
 
 file(GLOB headers xbyak/*.h)
 
diff --git a/doc/changelog.md b/doc/changelog.md
index aec29a4..df966c9 100644
--- a/doc/changelog.md
+++ b/doc/changelog.md
@@ -1,5 +1,6 @@
 # History
 
+* 2023/Dec/26 ver 7.03 set the default value of dfv to 0
 * 2023/Dec/20 ver 7.02 SHA* support APX
 * 2023/Dec/19 ver 7.01 support AESKLE, WIDE_KL, KEYLOCKER, KEYLOCKER_WIDE, detection of APX10/APX
 * 2023/Dec/01 ver 7.00 support APX
diff --git a/doc/usage.md b/doc/usage.md
index e0ecccc..0911b91 100644
--- a/doc/usage.md
+++ b/doc/usage.md
@@ -145,6 +145,16 @@ vpdpbusd(xm0, xm1, xm2); // VEX encoding
   - `imul(ax|T_zu, cx, 0x1234);` // Set ND=ZU
   - `imul(ax|T_zu|T_nf, cx, 0x1234);` // Set ND=ZU and EVEX.NF=1
   - `setb(r31b|T_zu);` // same as set(r31b); movzx(r31, r31b);
+  - See [sample/zero_upper.cpp](../sample/zero_upper.cpp)
+
+### ccmpSCC and ctestSCC
+
+- ccmpSCC(op1, op2, dfv = 0); // eflags = eflags == SCC ? cmp(op1, op2) : dfv
+- ctestSCC(op1, op2, dfv = 0); // eflags = eflags == SCC ? test(op1, op2) : dfv
+- SCC means source condition code such as z, a, gt.
+- See [sample/ccmp.cpp](../sample/ccmp.cpp)
+- Specify the union of T_of(=8), T_sf(=4), T_zf(=2), or T_cf(=1) for dfv.
+
 
 ## Label
 Two kinds of Label are supported. (String literal and Label class).
diff --git a/gen/gen_code.cpp b/gen/gen_code.cpp
index d1a0bd7..13b9ab6 100644
--- a/gen/gen_code.cpp
+++ b/gen/gen_code.cpp
@@ -635,10 +635,10 @@ void put()
 			// ccmpscc
 			// true if SCC = 0b1010, false if SCC = 0b1011 (see APX Architecture Specification p.266)
 			const char *s = p->ext == 10 ? "t" : p->ext == 11 ? "f" : p->name;
-			printf("void ccmp%s(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, %d); }\n", s, p->ext);
-			printf("void ccmp%s(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, %d); }\n", s, p->ext);
-			printf("void ctest%s(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, %d); }\n", s, p->ext);
-			printf("void ctest%s(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, %d); }\n", s, p->ext);
+			printf("void ccmp%s(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, %d); }\n", s, p->ext);
+			printf("void ccmp%s(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, %d); }\n", s, p->ext);
+			printf("void ctest%s(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, %d); }\n", s, p->ext);
+			printf("void ctest%s(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, %d); }\n", s, p->ext);
 		}
 	}
 	{
diff --git a/meson.build b/meson.build
index edc97cd..abe46f8 100644
--- a/meson.build
+++ b/meson.build
@@ -5,7 +5,7 @@
 project(
 	'xbyak',
 	'cpp',
-	version: '7.02',
+	version: '7.03',
 	license: 'BSD-3-Clause',
 	default_options: 'b_ndebug=if-release'
 )
diff --git a/readme.md b/readme.md
index 0de9024..42aedf6 100644
--- a/readme.md
+++ b/readme.md
@@ -1,5 +1,5 @@
 
-# Xbyak 7.02 [![Badge Build]][Build Status]
+# Xbyak 7.03 [![Badge Build]][Build Status]
 
 *A C++ JIT assembler for x86 (IA32), x64 (AMD64, x86-64)*
 
diff --git a/readme.txt b/readme.txt
index 08e9deb..6b5b4bb 100644
--- a/readme.txt
+++ b/readme.txt
@@ -1,5 +1,5 @@
 

-    C++用x86(IA-32), x64(AMD64, x86-64) JITアセンブラ Xbyak 7.02

+    C++用x86(IA-32), x64(AMD64, x86-64) JITアセンブラ Xbyak 7.03

 

 -----------------------------------------------------------------------------

 ◎概要

@@ -404,6 +404,7 @@ sample/{echo,hello}.bfは http://www.kmonos.net/alang/etc/brainfuck.php から
 -----------------------------------------------------------------------------

 ◎履歴

 

+2023/12/26 ver 7.03 dfvのデフォルト値を0に設定

 2023/12/20 ver 7.02 SHA*のAPX対応

 2023/12/19 ver 7.01 AESKLE, WIDE_KL, KEYLOCKER, KEYLOCKER_WIDE対応 APX10/APX判定対応

 2023/12/01 ver 7.00 APX対応

diff --git a/sample/Makefile b/sample/Makefile
index 62c7dcf..2010f69 100644
--- a/sample/Makefile
+++ b/sample/Makefile
@@ -30,7 +30,7 @@ else
 endif
 
 ifeq ($(BIT),64)
-TARGET += test64 bf64 memfunc64 test_util64 jmp_table64
+TARGET += test64 bf64 memfunc64 test_util64 jmp_table64 zero_upper ccmp
 ifeq ($(BOOST_EXIST),1)
 TARGET += calc64 #calc2_64
 endif
@@ -103,6 +103,14 @@ profiler: profiler.cpp ../xbyak/xbyak_util.h
 	$(CXX) $(CFLAGS) profiler.cpp -o $@
 profiler-vtune: profiler.cpp ../xbyak/xbyak_util.h
 	$(CXX) $(CFLAGS) profiler.cpp -o $@ -DXBYAK_USE_VTUNE -I /opt/intel/vtune_amplifier/include/ -L /opt/intel/vtune_amplifier/lib64 -ljitprofiling -ldl
+zero_upper: zero_upper.cpp $(XBYAK_INC)
+	$(CXX) $(CFLAGS) zero_upper.cpp -o $@
+test_zero_upper: zero_upper
+	sde -future -- ./zero_upper
+ccmp: ccmp.cpp $(XBYAK_INC)
+	$(CXX) $(CFLAGS) ccmp.cpp -o $@
+test_ccmp: ccmp
+	sde -future -- ./ccmp
 
 clean:
 	rm -rf $(TARGET) profiler profiler-vtune
diff --git a/sample/ccmp.cpp b/sample/ccmp.cpp
new file mode 100644
index 0000000..622d3d1
--- /dev/null
+++ b/sample/ccmp.cpp
@@ -0,0 +1,68 @@
+/*
+	An example of ccmp
+	> g++ ccmp.cpp -I ../xbyak
+	> sde -future -- ./a.out
+*/
+#include <stdio.h>
+#include <xbyak/xbyak.h>
+#include <xbyak/xbyak_util.h>
+
+using namespace Xbyak;
+
+struct Code1 : Xbyak::CodeGenerator {
+	Code1()
+	{
+		Xbyak::util::StackFrame sf(this, 2);
+		const auto& p1 = sf.p[0];
+		const auto& p2 = sf.p[1];
+		int dfv = 0;
+		cmp(p1, 3);
+		ctesta(p2, 1, dfv); // eflags = (p1 > 3) ? (p2 & 1) : dfv;
+		setz(al|T_zu);
+	}
+};
+
+struct Code2 : Xbyak::CodeGenerator {
+	Code2()
+	{
+		Xbyak::util::StackFrame sf(this, 3);
+		const auto& p1 = sf.p[0];
+		const auto& p2 = sf.p[1];
+		const auto& p3 = sf.p[2];
+		int dfv = 0;
+		cmp(p1, 1);
+		ccmpe(p2, 2, dfv); // eflags = p1==1 ? p2==2 : dfv;
+		ccmpe(p3, 3, dfv); // eflags = (p1==1 && p2==2) ? p3==3 : dfv;
+		setz(al|T_zu); // p1==1 && p2==2 && p3==3
+	}
+};
+
+
+int main()
+	try
+{
+	{
+		puts("(p1 > 3) && ((p2 & 1) == 0)");
+		Code1 c;
+		auto f = c.getCode<int (*)(int, int)>();
+		for (int p1 = 2; p1 < 5; p1++) {
+			for (int p2 = 0; p2 < 3; p2++) {
+				printf("p1=%d p2=%d ret=%d (%d)\n", p1, p2, f(p1, p2), p1 > 3 && ((p2&1) == 0));
+			}
+		}
+	}
+	{
+		puts("p1 == 1 && p2 == 2 && p3 == 3");
+		Code2 c;
+		auto f = c.getCode<int (*)(int, int, int)>();
+		for (int p1 = 0; p1 < 3; p1++) {
+			for (int p2 = 1; p2 < 4; p2++) {
+				for (int p3 = 2; p3 < 5; p3++) {
+					printf("p1=%d p2=%d p3=%d ret=%d (%d)\n", p1, p2, p3, f(p1, p2, p3), p1==1 && p2==2 && p3==3);
+				}
+			}
+		}
+	}
+} catch (std::exception& e) {
+	printf("ERR %s\n", e.what());
+}
diff --git a/sample/zero_upper.cpp b/sample/zero_upper.cpp
new file mode 100644
index 0000000..46dcca9
--- /dev/null
+++ b/sample/zero_upper.cpp
@@ -0,0 +1,48 @@
+/*
+	An example of T_zu (zero upper) flag
+	> g++ zero_upper.cpp -I ../xbyak
+	> sde -future -- ./a.out
+*/
+#include <stdio.h>
+#include <xbyak/xbyak.h>
+
+using namespace Xbyak;
+
+struct Code : Xbyak::CodeGenerator {
+	Code(int mode)
+	{
+		mov(eax, 0x12345678);
+		cmp(eax, eax); // ZF=1
+		switch (mode) {
+		case 0: // imul
+			puts("imul");
+			imul(ax,ax, 0x1234);
+			break;
+		case 1: // imul+zu
+			puts("imul+zu");
+			imul(ax|T_zu, ax, 0x1234);
+			break;
+		case 2: // setz
+			puts("setz");
+			setz(al);
+			break;
+		case 3: // setz+zu
+			puts("setz+zu");
+			setz(al|T_zu);
+			break;
+		}
+		ret();
+	}
+};
+
+int main()
+	try
+{
+	for (int mode = 0; mode < 4; mode++) {
+		Code c(mode);
+		auto f = c.getCode<int (*)()>();
+		printf("ret=%08x\n", f());
+	}
+} catch (std::exception& e) {
+	printf("ERR %s\n", e.what());
+}
diff --git a/xbyak/xbyak.h b/xbyak/xbyak.h
index c99e809..7c20b13 100644
--- a/xbyak/xbyak.h
+++ b/xbyak/xbyak.h
@@ -155,7 +155,7 @@ namespace Xbyak {
 
 enum {
 	DEFAULT_MAX_CODE_SIZE = 4096,
-	VERSION = 0x7020 /* 0xABCD = A.BC(.D) */
+	VERSION = 0x7030 /* 0xABCD = A.BC(.D) */
 };
 
 #ifndef MIE_INTEGER_TYPE_DEFINED
diff --git a/xbyak/xbyak_mnemonic.h b/xbyak/xbyak_mnemonic.h
index d863d46..5392492 100644
--- a/xbyak/xbyak_mnemonic.h
+++ b/xbyak/xbyak_mnemonic.h
@@ -1,4 +1,4 @@
-const char *getVersionString() const { return "7.02"; }
+const char *getVersionString() const { return "7.03"; }
 void aadd(const Address& addr, const Reg32e &reg) { opMR(addr, reg, T_0F38, 0x0FC); }
 void aand(const Address& addr, const Reg32e &reg) { opMR(addr, reg, T_0F38 | T_66, 0x0FC); }
 void adc(const Operand& op, uint32_t imm) { opOI(op, imm, 0x10, 2); }
@@ -66,62 +66,62 @@ void bts(const Operand& op, const Reg& reg) { opRO(reg, op, T_0F, 0xAB, op.isREG
 void bts(const Operand& op, uint8_t imm) { opRext(op, 16|i32e, 5, T_0F, 0xba, false, 1); db(imm); }
 void bzhi(const Reg32e& r1, const Operand& op, const Reg32e& r2) { opRRO(r1, r2, op, T_APX|T_0F38|T_NF, 0xf5); }
 void cbw() { db(0x66); db(0x98); }
-void ccmpa(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 7); }
-void ccmpa(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 7); }
-void ccmpae(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 3); }
-void ccmpae(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 3); }
-void ccmpb(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 2); }
-void ccmpb(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 2); }
-void ccmpbe(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 6); }
-void ccmpbe(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 6); }
-void ccmpc(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 2); }
-void ccmpc(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 2); }
-void ccmpe(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 4); }
-void ccmpe(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 4); }
-void ccmpf(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 11); }
-void ccmpf(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 11); }
-void ccmpg(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 15); }
-void ccmpg(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 15); }
-void ccmpge(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 13); }
-void ccmpge(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 13); }
-void ccmpl(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 12); }
-void ccmpl(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 12); }
-void ccmple(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 14); }
-void ccmple(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 14); }
-void ccmpna(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 6); }
-void ccmpna(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 6); }
-void ccmpnae(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 2); }
-void ccmpnae(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 2); }
-void ccmpnb(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 3); }
-void ccmpnb(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 3); }
-void ccmpnbe(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 7); }
-void ccmpnbe(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 7); }
-void ccmpnc(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 3); }
-void ccmpnc(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 3); }
-void ccmpne(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 5); }
-void ccmpne(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 5); }
-void ccmpng(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 14); }
-void ccmpng(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 14); }
-void ccmpnge(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 12); }
-void ccmpnge(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 12); }
-void ccmpnl(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 13); }
-void ccmpnl(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 13); }
-void ccmpnle(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 15); }
-void ccmpnle(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 15); }
-void ccmpno(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 1); }
-void ccmpno(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 1); }
-void ccmpns(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 9); }
-void ccmpns(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 9); }
-void ccmpnz(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 5); }
-void ccmpnz(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 5); }
-void ccmpo(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 0); }
-void ccmpo(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 0); }
-void ccmps(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 8); }
-void ccmps(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 8); }
-void ccmpt(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 10); }
-void ccmpt(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 10); }
-void ccmpz(const Operand& op, int imm, int dfv) { opCcmpi(op, imm, dfv, 4); }
-void ccmpz(const Operand& op1, const Operand& op2, int dfv) { opCcmp(op1, op2, dfv, 0x38, 4); }
+void ccmpa(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 7); }
+void ccmpa(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 7); }
+void ccmpae(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 3); }
+void ccmpae(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 3); }
+void ccmpb(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 2); }
+void ccmpb(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 2); }
+void ccmpbe(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 6); }
+void ccmpbe(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 6); }
+void ccmpc(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 2); }
+void ccmpc(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 2); }
+void ccmpe(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 4); }
+void ccmpe(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 4); }
+void ccmpf(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 11); }
+void ccmpf(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 11); }
+void ccmpg(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 15); }
+void ccmpg(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 15); }
+void ccmpge(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 13); }
+void ccmpge(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 13); }
+void ccmpl(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 12); }
+void ccmpl(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 12); }
+void ccmple(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 14); }
+void ccmple(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 14); }
+void ccmpna(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 6); }
+void ccmpna(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 6); }
+void ccmpnae(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 2); }
+void ccmpnae(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 2); }
+void ccmpnb(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 3); }
+void ccmpnb(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 3); }
+void ccmpnbe(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 7); }
+void ccmpnbe(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 7); }
+void ccmpnc(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 3); }
+void ccmpnc(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 3); }
+void ccmpne(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 5); }
+void ccmpne(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 5); }
+void ccmpng(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 14); }
+void ccmpng(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 14); }
+void ccmpnge(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 12); }
+void ccmpnge(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 12); }
+void ccmpnl(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 13); }
+void ccmpnl(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 13); }
+void ccmpnle(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 15); }
+void ccmpnle(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 15); }
+void ccmpno(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 1); }
+void ccmpno(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 1); }
+void ccmpns(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 9); }
+void ccmpns(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 9); }
+void ccmpnz(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 5); }
+void ccmpnz(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 5); }
+void ccmpo(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 0); }
+void ccmpo(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 0); }
+void ccmps(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 8); }
+void ccmps(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 8); }
+void ccmpt(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 10); }
+void ccmpt(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 10); }
+void ccmpz(const Operand& op, int imm, int dfv = 0) { opCcmpi(op, imm, dfv, 4); }
+void ccmpz(const Operand& op1, const Operand& op2, int dfv = 0) { opCcmp(op1, op2, dfv, 0x38, 4); }
 void cdq() { db(0x99); }
 void cfcmovb(const Operand& op1, const Operand& op2) { opCfcmov(Reg(), op1, op2, 0x42); }
 void cfcmovb(const Reg& d, const Reg& r, const Operand& op) { opCfcmov(d|T_nf, op, r, 0x42); }
@@ -271,62 +271,62 @@ void comisd(const Xmm& xmm, const Operand& op) { opSSE(xmm, op, T_66|T_0F, 0x2F,
 void comiss(const Xmm& xmm, const Operand& op) { opSSE(xmm, op, T_0F, 0x2F, isXMM_XMMorMEM); }
 void cpuid() { db(0x0F); db(0xA2); }
 void crc32(const Reg32e& r, const Operand& op) { if (!((r.isBit(32) && op.isBit(8|16|32)) || (r.isBit(64) && op.isBit(8|64)))) XBYAK_THROW(ERR_BAD_SIZE_OF_REGISTER) int code = 0xF0 | (op.isBit(8) ? 0 : 1); uint64_t type = op.isBit(16) ? T_66:0; if (opROO(Reg(), op, static_cast<const Reg&>(r), T_APX|type, code)) return; opRO(r, op, T_F2|T_0F38|type, code); }
-void ctesta(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 7); }
-void ctesta(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 7); }
-void ctestae(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 3); }
-void ctestae(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 3); }
-void ctestb(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 2); }
-void ctestb(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 2); }
-void ctestbe(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 6); }
-void ctestbe(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 6); }
-void ctestc(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 2); }
-void ctestc(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 2); }
-void cteste(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 4); }
-void cteste(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 4); }
-void ctestf(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 11); }
-void ctestf(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 11); }
-void ctestg(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 15); }
-void ctestg(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 15); }
-void ctestge(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 13); }
-void ctestge(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 13); }
-void ctestl(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 12); }
-void ctestl(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 12); }
-void ctestle(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 14); }
-void ctestle(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 14); }
-void ctestna(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 6); }
-void ctestna(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 6); }
-void ctestnae(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 2); }
-void ctestnae(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 2); }
-void ctestnb(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 3); }
-void ctestnb(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 3); }
-void ctestnbe(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 7); }
-void ctestnbe(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 7); }
-void ctestnc(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 3); }
-void ctestnc(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 3); }
-void ctestne(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 5); }
-void ctestne(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 5); }
-void ctestng(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 14); }
-void ctestng(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 14); }
-void ctestnge(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 12); }
-void ctestnge(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 12); }
-void ctestnl(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 13); }
-void ctestnl(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 13); }
-void ctestnle(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 15); }
-void ctestnle(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 15); }
-void ctestno(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 1); }
-void ctestno(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 1); }
-void ctestns(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 9); }
-void ctestns(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 9); }
-void ctestnz(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 5); }
-void ctestnz(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 5); }
-void ctesto(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 0); }
-void ctesto(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 0); }
-void ctests(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 8); }
-void ctests(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 8); }
-void ctestt(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 10); }
-void ctestt(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 10); }
-void ctestz(const Operand& op, const Reg& r, int dfv) { opCcmp(op, r, dfv, 0x84, 4); }
-void ctestz(const Operand& op, int imm, int dfv) { opTesti(op, imm, dfv, 4); }
+void ctesta(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 7); }
+void ctesta(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 7); }
+void ctestae(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 3); }
+void ctestae(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 3); }
+void ctestb(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 2); }
+void ctestb(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 2); }
+void ctestbe(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 6); }
+void ctestbe(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 6); }
+void ctestc(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 2); }
+void ctestc(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 2); }
+void cteste(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 4); }
+void cteste(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 4); }
+void ctestf(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 11); }
+void ctestf(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 11); }
+void ctestg(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 15); }
+void ctestg(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 15); }
+void ctestge(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 13); }
+void ctestge(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 13); }
+void ctestl(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 12); }
+void ctestl(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 12); }
+void ctestle(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 14); }
+void ctestle(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 14); }
+void ctestna(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 6); }
+void ctestna(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 6); }
+void ctestnae(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 2); }
+void ctestnae(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 2); }
+void ctestnb(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 3); }
+void ctestnb(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 3); }
+void ctestnbe(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 7); }
+void ctestnbe(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 7); }
+void ctestnc(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 3); }
+void ctestnc(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 3); }
+void ctestne(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 5); }
+void ctestne(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 5); }
+void ctestng(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 14); }
+void ctestng(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 14); }
+void ctestnge(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 12); }
+void ctestnge(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 12); }
+void ctestnl(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 13); }
+void ctestnl(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 13); }
+void ctestnle(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 15); }
+void ctestnle(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 15); }
+void ctestno(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 1); }
+void ctestno(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 1); }
+void ctestns(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 9); }
+void ctestns(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 9); }
+void ctestnz(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 5); }
+void ctestnz(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 5); }
+void ctesto(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 0); }
+void ctesto(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 0); }
+void ctests(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 8); }
+void ctests(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 8); }
+void ctestt(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 10); }
+void ctestt(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 10); }
+void ctestz(const Operand& op, const Reg& r, int dfv = 0) { opCcmp(op, r, dfv, 0x84, 4); }
+void ctestz(const Operand& op, int imm, int dfv = 0) { opTesti(op, imm, dfv, 4); }
 void cvtdq2pd(const Xmm& xmm, const Operand& op) { opSSE(xmm, op, T_F3|T_0F, 0xE6, isXMM_XMMorMEM); }
 void cvtdq2ps(const Xmm& xmm, const Operand& op) { opSSE(xmm, op, T_0F, 0x5B, isXMM_XMMorMEM); }
 void cvtpd2dq(const Xmm& xmm, const Operand& op) { opSSE(xmm, op, T_F2|T_0F, 0xE6, isXMM_XMMorMEM); }