diff --git a/gen/gen_code.cpp b/gen/gen_code.cpp
index cc4e005..e8abcc5 100644
--- a/gen/gen_code.cpp
+++ b/gen/gen_code.cpp
@@ -1034,8 +1034,8 @@ void put()
 			{ 0x0C, "blendps", MM_0F3A | PP_66, true, 0, true, true },
 			{ 0x41, "dppd", MM_0F3A | PP_66, false, 0, true, true },
 			{ 0x40, "dpps", MM_0F3A | PP_66, true, 0, true, true },
-			{ 0x42, "mpsadbw", MM_0F3A | PP_66, false, 0, true, true },
-			{ 0x0E, "pblendw", MM_0F3A | PP_66, false, 0, true, true },
+			{ 0x42, "mpsadbw", MM_0F3A | PP_66, true, 0, true, true },
+			{ 0x0E, "pblendw", MM_0F3A | PP_66, true, 0, true, true },
 			{ 0x0B, "roundsd", MM_0F3A | PP_66, false, 0, true, true },
 			{ 0x0A, "roundss", MM_0F3A | PP_66, false, 0, true, true },
 			{ 0x44, "pclmulqdq", MM_0F3A | PP_66, false, 0, true, true },
@@ -1220,12 +1220,12 @@ void put()
 			{ 0x10, "movupd", MM_0F | PP_66, true, -1, false },
 			{ 0x10, "movups", MM_0F, true, -1, false },
 
-			{ 0x1C, "pabsb", MM_0F38 | PP_66, false, -1, false },
-			{ 0x1D, "pabsw", MM_0F38 | PP_66, false, -1, false },
-			{ 0x1E, "pabsd", MM_0F38 | PP_66, false, -1, false },
+			{ 0x1C, "pabsb", MM_0F38 | PP_66, true, -1, false },
+			{ 0x1D, "pabsw", MM_0F38 | PP_66, true, -1, false },
+			{ 0x1E, "pabsd", MM_0F38 | PP_66, true, -1, false },
 			{ 0x41, "phminposuw", MM_0F38 | PP_66, false, -1, false },
 
-			{ 0x20, "pmovsxbw", MM_0F38 | PP_66, false, -1, false },
+			{ 0x20, "pmovsxbw", MM_0F38 | PP_66, true, -1, false },
 			{ 0x21, "pmovsxbd", MM_0F38 | PP_66, false, -1, false },
 			{ 0x22, "pmovsxbq", MM_0F38 | PP_66, false, -1, false },
 			{ 0x23, "pmovsxwd", MM_0F38 | PP_66, false, -1, false },
@@ -1492,7 +1492,7 @@ void put()
 		printf("void vmovntdq(const Address& addr, const Xmm& x) { opAVX_X_X_XM(x, x.isXMM() ? xm0 : ym0, addr, MM_0F | PP_66, 0xE7, true); }\n");
 		printf("void vmovntpd(const Address& addr, const Xmm& x) { opAVX_X_X_XM(x, x.isXMM() ? xm0 : ym0, addr, MM_0F | PP_66, 0x2B, true); }\n");
 		printf("void vmovntps(const Address& addr, const Xmm& x) { opAVX_X_X_XM(x, x.isXMM() ? xm0 : ym0, addr, MM_0F, 0x2B, true); }\n");
-		printf("void vmovntdqa(const Xmm& x, const Address& addr) { opAVX_X_X_XM(x, xm0, addr, MM_0F38 | PP_66, 0x2A, false); }\n");
+		printf("void vmovntdqa(const Xmm& x, const Address& addr) { opAVX_X_X_XM(x, x.isXMM() ? xm0 : ymm0, addr, MM_0F38 | PP_66, 0x2A, true); }\n");
 
 		// vmovsd, vmovss
 		for (int i = 0; i < 2; i++) {
diff --git a/test/make_nm.cpp b/test/make_nm.cpp
index feaf901..30b8374 100644
--- a/test/make_nm.cpp
+++ b/test/make_nm.cpp
@@ -1413,8 +1413,8 @@ class Test {
 			{ "vblendps", true },
 			{ "vdppd", false },
 			{ "vdpps", true },
-			{ "vmpsadbw", false },
-			{ "vpblendw", false },
+			{ "vmpsadbw", true },
+			{ "vpblendw", true },
 			{ "vroundsd", false },
 			{ "vroundss", false },
 			{ "vpclmulqdq", false },
@@ -1501,9 +1501,9 @@ class Test {
 			{ "vmovupd", true },
 			{ "vmovups", true },
 
-			{ "vpabsb", false },
-			{ "vpabsw", false },
-			{ "vpabsd", false },
+			{ "vpabsb", true },
+			{ "vpabsw", true },
+			{ "vpabsd", true },
 			{ "vphminposuw", false },
 
 			{ "vpmovsxbw", false },
@@ -1539,6 +1539,30 @@ class Test {
 			put(p->name, YMM, YMM | MEM);
 		}
 	}
+	void putAVX_Y_XM()
+	{
+		const char *tbl[] = {
+			"vpmovsxbw",
+#if 0
+			"vpmovsxbd",
+			"vpmovsxbq",
+			"vpmovsxwd",
+			"vpmovsxwq",
+			"vpmovsxdq",
+
+			"vpmovzxbw",
+			"vpmovzxbd",
+			"vpmovzxbq",
+			"vpmovzxwd",
+			"vpmovzxwq",
+			"vpmovzxdq",
+#endif
+		};
+		for (size_t i = 0; i < NUM_OF_ARRAY(tbl); i++) {
+			const char *name = tbl[i];
+			put(name, YMM, XMM);
+		}
+	}
 	void putAVX_M_X()
 	{
 		const struct Tbl {
@@ -1659,7 +1683,7 @@ class Test {
 		put("vmovntdq", MEM, XMM | YMM);
 		put("vmovntpd", MEM, XMM | YMM);
 		put("vmovntps", MEM, XMM | YMM);
-		put("vmovntdqa", XMM, MEM);
+		put("vmovntdqa", XMM | YMM, MEM);
 
 		{
 			const char tbl[][8] = { "vmovsd", "vmovss" };
@@ -1918,6 +1942,7 @@ public:
 		putAVX_X_XM();
 		putAVX_M_X();
 		putAVX_X_X_IMM_omit();
+		putAVX_Y_XM();
 		putFMA();
 #endif
 #else // USE_AVX
diff --git a/xbyak/xbyak_mnemonic.h b/xbyak/xbyak_mnemonic.h
index 2898742..6226d47 100644
--- a/xbyak/xbyak_mnemonic.h
+++ b/xbyak/xbyak_mnemonic.h
@@ -641,10 +641,10 @@ void vdppd(const Xmm& xm1, const Xmm& xm2, const Operand& op, uint8 imm) { opAVX
 void vdppd(const Xmm& xmm, const Operand& op, uint8 imm) { opAVX_X_X_XM(xmm, xmm, op, MM_0F3A | PP_66, 0x41, false, 0); db(imm); }
 void vdpps(const Xmm& xm1, const Xmm& xm2, const Operand& op, uint8 imm) { opAVX_X_X_XM(xm1, xm2, op, MM_0F3A | PP_66, 0x40, true, 0); db(imm); }
 void vdpps(const Xmm& xmm, const Operand& op, uint8 imm) { opAVX_X_X_XM(xmm, xmm, op, MM_0F3A | PP_66, 0x40, true, 0); db(imm); }
-void vmpsadbw(const Xmm& xm1, const Xmm& xm2, const Operand& op, uint8 imm) { opAVX_X_X_XM(xm1, xm2, op, MM_0F3A | PP_66, 0x42, false, 0); db(imm); }
-void vmpsadbw(const Xmm& xmm, const Operand& op, uint8 imm) { opAVX_X_X_XM(xmm, xmm, op, MM_0F3A | PP_66, 0x42, false, 0); db(imm); }
-void vpblendw(const Xmm& xm1, const Xmm& xm2, const Operand& op, uint8 imm) { opAVX_X_X_XM(xm1, xm2, op, MM_0F3A | PP_66, 0x0E, false, 0); db(imm); }
-void vpblendw(const Xmm& xmm, const Operand& op, uint8 imm) { opAVX_X_X_XM(xmm, xmm, op, MM_0F3A | PP_66, 0x0E, false, 0); db(imm); }
+void vmpsadbw(const Xmm& xm1, const Xmm& xm2, const Operand& op, uint8 imm) { opAVX_X_X_XM(xm1, xm2, op, MM_0F3A | PP_66, 0x42, true, 0); db(imm); }
+void vmpsadbw(const Xmm& xmm, const Operand& op, uint8 imm) { opAVX_X_X_XM(xmm, xmm, op, MM_0F3A | PP_66, 0x42, true, 0); db(imm); }
+void vpblendw(const Xmm& xm1, const Xmm& xm2, const Operand& op, uint8 imm) { opAVX_X_X_XM(xm1, xm2, op, MM_0F3A | PP_66, 0x0E, true, 0); db(imm); }
+void vpblendw(const Xmm& xmm, const Operand& op, uint8 imm) { opAVX_X_X_XM(xmm, xmm, op, MM_0F3A | PP_66, 0x0E, true, 0); db(imm); }
 void vroundsd(const Xmm& xm1, const Xmm& xm2, const Operand& op, uint8 imm) { opAVX_X_X_XM(xm1, xm2, op, MM_0F3A | PP_66, 0x0B, false, 0); db(imm); }
 void vroundsd(const Xmm& xmm, const Operand& op, uint8 imm) { opAVX_X_X_XM(xmm, xmm, op, MM_0F3A | PP_66, 0x0B, false, 0); db(imm); }
 void vroundss(const Xmm& xm1, const Xmm& xm2, const Operand& op, uint8 imm) { opAVX_X_X_XM(xm1, xm2, op, MM_0F3A | PP_66, 0x0A, false, 0); db(imm); }
@@ -877,11 +877,11 @@ void vmovshdup(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F
 void vmovsldup(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F | PP_F3, 0x12, true, -1); }
 void vmovupd(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F | PP_66, 0x10, true, -1); }
 void vmovups(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F, 0x10, true, -1); }
-void vpabsb(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F38 | PP_66, 0x1C, false, -1); }
-void vpabsw(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F38 | PP_66, 0x1D, false, -1); }
-void vpabsd(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F38 | PP_66, 0x1E, false, -1); }
+void vpabsb(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F38 | PP_66, 0x1C, true, -1); }
+void vpabsw(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F38 | PP_66, 0x1D, true, -1); }
+void vpabsd(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F38 | PP_66, 0x1E, true, -1); }
 void vphminposuw(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F38 | PP_66, 0x41, false, -1); }
-void vpmovsxbw(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F38 | PP_66, 0x20, false, -1); }
+void vpmovsxbw(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F38 | PP_66, 0x20, true, -1); }
 void vpmovsxbd(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F38 | PP_66, 0x21, false, -1); }
 void vpmovsxbq(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F38 | PP_66, 0x22, false, -1); }
 void vpmovsxwd(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, MM_0F38 | PP_66, 0x23, false, -1); }
@@ -1342,7 +1342,7 @@ void vmovmskps(const Reg& r, const Xmm& x) { if (!r.isBit(i32e)) throw ERR_BAD_C
 void vmovntdq(const Address& addr, const Xmm& x) { opAVX_X_X_XM(x, x.isXMM() ? xm0 : ym0, addr, MM_0F | PP_66, 0xE7, true); }
 void vmovntpd(const Address& addr, const Xmm& x) { opAVX_X_X_XM(x, x.isXMM() ? xm0 : ym0, addr, MM_0F | PP_66, 0x2B, true); }
 void vmovntps(const Address& addr, const Xmm& x) { opAVX_X_X_XM(x, x.isXMM() ? xm0 : ym0, addr, MM_0F, 0x2B, true); }
-void vmovntdqa(const Xmm& x, const Address& addr) { opAVX_X_X_XM(x, xm0, addr, MM_0F38 | PP_66, 0x2A, false); }
+void vmovntdqa(const Xmm& x, const Address& addr) { opAVX_X_X_XM(x, x.isXMM() ? xm0 : ymm0, addr, MM_0F38 | PP_66, 0x2A, true); }
 void vmovsd(const Xmm& x1, const Xmm& x2, const Operand& op = Operand()) { if (!op.isNone() && !op.isXMM()) throw ERR_BAD_COMBINATION; opAVX_X_X_XM(x1, x2, op, MM_0F | PP_F2, 0x10, false); }
 void vmovsd(const Xmm& x, const Address& addr) { opAVX_X_X_XM(x, xm0, addr, MM_0F | PP_F2, 0x10, false); }
 void vmovsd(const Address& addr, const Xmm& x) { opAVX_X_X_XM(x, xm0, addr, MM_0F | PP_F2, 0x11, false); }