Update Yasm to 1.2.0 (Yasm Part 1/3)

third_party/yasm/patched-yasm/x86insn_nasm.gperf

Index: third_party/yasm/patched-yasm/x86insn_nasm.gperf
===================================================================
--- third_party/yasm/patched-yasm/x86insn_nasm.gperf	(revision 165590)
+++ third_party/yasm/patched-yasm/x86insn_nasm.gperf	(working copy)
@@ -1,4 +1,4 @@
-/* Generated by gen_x86_insn.py r2346, do not edit */
+/* Generated by gen_x86_insn.py rHEAD, do not edit */
 %ignore-case
 %language=ANSI-C
 %compare-strncmp
@@ -31,15 +31,20 @@
 aesimc,	aesimc_insn,	1,	SUF_Z,	0x38,	0xDB,	0,	0,	CPU_AES,	0,	0
 aeskeygenassist,	aes_imm_insn,	1,	SUF_Z,	0x3A,	0xDF,	0,	0,	CPU_AES,	0,	0
 and,	arith_insn,	22,	SUF_Z,	0x20,	0x04,	0,	0,	0,	0,	0
+andn,	vex_gpr_reg_nds_rm_0F_insn,	2,	SUF_Z,	0x00,	0x38,	0xF2,	ONLY_AVX,	CPU_BMI1,	0,	0
 andnpd,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x55,	0,	0,	CPU_SSE2,	0,	0
 andnps,	xmm_xmm128_insn,	2,	SUF_Z,	0x00,	0x55,	0,	0,	CPU_SSE,	0,	0
 andpd,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x54,	0,	0,	CPU_SSE2,	0,	0
 andps,	xmm_xmm128_insn,	2,	SUF_Z,	0x00,	0x54,	0,	0,	CPU_SSE,	0,	0
 arpl,	arpl_insn,	1,	SUF_Z,	0,	0,	0,	NOT_64,	CPU_286,	CPU_Prot,	0
+bextr,	vex_gpr_reg_rm_nds_0F_insn,	2,	SUF_Z,	0x00,	0x38,	0xF7,	ONLY_AVX,	CPU_BMI1,	0,	0
 blendpd,	sse4imm_insn,	2,	SUF_Z,	0x0D,	0,	0,	0,	CPU_SSE41,	0,	0
 blendps,	sse4imm_insn,	2,	SUF_Z,	0x0C,	0,	0,	0,	CPU_SSE41,	0,	0
 blendvpd,	sse4xmm0_insn,	2,	SUF_Z,	0x15,	0,	0,	0,	CPU_SSE41,	0,	0
 blendvps,	sse4xmm0_insn,	2,	SUF_Z,	0x14,	0,	0,	0,	CPU_SSE41,	0,	0
+blsi,	vex_gpr_ndd_rm_0F38_regext_insn,	2,	SUF_Z,	0x00,	0xF3,	0x03,	ONLY_AVX,	CPU_BMI1,	0,	0
+blsmsk,	vex_gpr_ndd_rm_0F38_regext_insn,	2,	SUF_Z,	0x00,	0xF3,	0x02,	ONLY_AVX,	CPU_BMI1,	0,	0
+blsr,	vex_gpr_ndd_rm_0F38_regext_insn,	2,	SUF_Z,	0x00,	0xF3,	0x01,	ONLY_AVX,	CPU_BMI1,	0,	0
 bound,	bound_insn,	2,	SUF_Z,	0,	0,	0,	NOT_64,	CPU_186,	0,	0
 bsf,	bsfr_insn,	3,	SUF_Z,	0xBC,	0,	0,	0,	CPU_386,	0,	0
 bsr,	bsfr_insn,	3,	SUF_Z,	0xBD,	0,	0,	0,	CPU_386,	0,	0
@@ -48,6 +53,7 @@
 btc,	bittest_insn,	6,	SUF_Z,	0xBB,	0x07,	0,	0,	CPU_386,	0,	0
 btr,	bittest_insn,	6,	SUF_Z,	0xB3,	0x06,	0,	0,	CPU_386,	0,	0
 bts,	bittest_insn,	6,	SUF_Z,	0xAB,	0x05,	0,	0,	CPU_386,	0,	0
+bzhi,	vex_gpr_reg_rm_nds_0F_insn,	2,	SUF_Z,	0x00,	0x38,	0xF5,	ONLY_AVX,	CPU_BMI2,	0,	0
 call,	call_insn,	30,	SUF_Z,	0,	0,	0,	0,	0,	0,	0
 cbw,	onebyte_insn,	1,	SUF_Z,	0x98,	0x10,	0,	0,	0,	0,	0
 cdq,	onebyte_insn,	1,	SUF_Z,	0x99,	0x20,	0,	0,	CPU_386,	0,	0
@@ -300,6 +306,7 @@
 invept,	eptvpid_insn,	2,	SUF_Z,	0x00,	0,	0,	0,	CPU_386,	CPU_EPTVPID,	0
 invlpg,	twobytemem_insn,	1,	SUF_Z,	0x07,	0x0F,	0x01,	0,	CPU_486,	CPU_Priv,	0
 invlpga,	invlpga_insn,	2,	SUF_Z,	0,	0,	0,	0,	CPU_SVM,	0,	0
+invpcid,	invpcid_insn,	2,	SUF_Z,	0,	0,	0,	0,	CPU_386,	CPU_INVPCID,	CPU_Priv
 invvpid,	eptvpid_insn,	2,	SUF_Z,	0x01,	0,	0,	0,	CPU_386,	CPU_EPTVPID,	0
 iret,	onebyte_insn,	1,	SUF_Z,	0xCF,	0,	0,	0,	0,	0,	0
 iretd,	onebyte_insn,	1,	SUF_Z,	0xCF,	0x20,	0,	0,	CPU_386,	0,	0
@@ -340,7 +347,7 @@
 js,	jcc_insn,	9,	SUF_Z,	0x08,	0,	0,	0,	0,	0,	0
 jz,	jcc_insn,	9,	SUF_Z,	0x04,	0,	0,	0,	0,	0,	0
 lahf,	onebyte_insn,	1,	SUF_Z,	0x9F,	0,	0,	0,	0,	0,	0
-lar,	bsfr_insn,	3,	SUF_Z,	0x02,	0,	0,	0,	CPU_286,	CPU_Prot,	0
+lar,	larlsl_insn,	6,	SUF_Z,	0x02,	0,	0,	0,	CPU_286,	CPU_Prot,	0
 lddqu,	lddqu_insn,	2,	SUF_Z,	0,	0,	0,	0,	CPU_SSE3,	0,	0
 ldmxcsr,	ldstmxcsr_insn,	1,	SUF_Z,	0x02,	0,	0,	0,	CPU_SSE,	0,	0
 lds,	ldes_insn,	2,	SUF_Z,	0xC5,	0,	0,	NOT_64,	0,	0,	0
@@ -348,9 +355,9 @@
 leave,	onebyte_insn,	1,	SUF_Z,	0xC9,	0x00,	0x40,	0,	CPU_186,	0,	0
 les,	ldes_insn,	2,	SUF_Z,	0xC4,	0,	0,	NOT_64,	0,	0,	0
 lfence,	threebyte_insn,	1,	SUF_Z,	0x0F,	0xAE,	0xE8,	0,	CPU_P3,	0,	0
-lfs,	lfgss_insn,	2,	SUF_Z,	0xB4,	0,	0,	0,	CPU_386,	0,	0
+lfs,	lfgss_insn,	3,	SUF_Z,	0xB4,	0,	0,	0,	CPU_386,	0,	0
 lgdt,	twobytemem_insn,	1,	SUF_Z,	0x02,	0x0F,	0x01,	0,	CPU_286,	CPU_Priv,	0
-lgs,	lfgss_insn,	2,	SUF_Z,	0xB5,	0,	0,	0,	CPU_386,	0,	0
+lgs,	lfgss_insn,	3,	SUF_Z,	0xB5,	0,	0,	0,	CPU_386,	0,	0
 lidt,	twobytemem_insn,	1,	SUF_Z,	0x03,	0x0F,	0x01,	0,	CPU_286,	CPU_Priv,	0
 lldt,	prot286_insn,	1,	SUF_Z,	0x02,	0x00,	0,	0,	CPU_286,	CPU_Priv,	CPU_Prot
 lmsw,	prot286_insn,	1,	SUF_Z,	0x06,	0x01,	0,	0,	CPU_286,	CPU_Priv,	0
@@ -366,10 +373,10 @@
 loopne,	loop_insn,	8,	SUF_Z,	0x00,	0,	0,	0,	0,	0,	0
 loopnz,	loop_insn,	8,	SUF_Z,	0x00,	0,	0,	0,	0,	0,	0
 loopz,	loop_insn,	8,	SUF_Z,	0x01,	0,	0,	0,	0,	0,	0
-lsl,	bsfr_insn,	3,	SUF_Z,	0x03,	0,	0,	0,	CPU_286,	CPU_Prot,	0
-lss,	lfgss_insn,	2,	SUF_Z,	0xB2,	0,	0,	0,	CPU_386,	0,	0
+lsl,	larlsl_insn,	6,	SUF_Z,	0x03,	0,	0,	0,	CPU_286,	CPU_Prot,	0
+lss,	lfgss_insn,	3,	SUF_Z,	0xB2,	0,	0,	0,	CPU_386,	0,	0
 ltr,	prot286_insn,	1,	SUF_Z,	0x03,	0x00,	0,	0,	CPU_286,	CPU_Priv,	CPU_Prot
-lzcnt,	cnt_insn,	3,	SUF_Z,	0xBD,	0,	0,	0,	CPU_686,	CPU_AMD,	0
+lzcnt,	cnt_insn,	3,	SUF_Z,	0xBD,	0,	0,	0,	CPU_LZCNT,	0,	0
 maskmovdqu,	maskmovdqu_insn,	1,	SUF_Z,	0,	0,	0,	0,	CPU_SSE2,	0,	0
 maskmovq,	maskmovq_insn,	1,	SUF_Z,	0,	0,	0,	0,	CPU_MMX,	CPU_P3,	0
 maxpd,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x5F,	0,	0,	CPU_SSE2,	0,	0
@@ -401,7 +408,7 @@
 movmskpd,	movmsk_insn,	4,	SUF_Z,	0x66,	0,	0,	0,	CPU_SSE2,	0,	0
 movmskps,	movmsk_insn,	4,	SUF_Z,	0,	0,	0,	0,	CPU_386,	CPU_SSE,	0
 movntdq,	movnt_insn,	2,	SUF_Z,	0x66,	0xE7,	0,	0,	CPU_SSE2,	0,	0
-movntdqa,	movntdqa_insn,	1,	SUF_Z,	0,	0,	0,	0,	CPU_SSE41,	0,	0
+movntdqa,	movntdqa_insn,	2,	SUF_Z,	0,	0,	0,	0,	CPU_SSE41,	0,	0
 movnti,	movnti_insn,	2,	SUF_Z,	0,	0,	0,	0,	CPU_P4,	0,	0
 movntpd,	movnt_insn,	2,	SUF_Z,	0x66,	0x2B,	0,	0,	CPU_SSE2,	0,	0
 movntps,	movnt_insn,	2,	SUF_Z,	0x00,	0x2B,	0,	0,	CPU_SSE,	0,	0
@@ -428,6 +435,7 @@
 mulps,	xmm_xmm128_insn,	2,	SUF_Z,	0x00,	0x59,	0,	0,	CPU_SSE,	0,	0
 mulsd,	xmm_xmm64_insn,	4,	SUF_Z,	0xF2,	0x59,	0,	0,	CPU_SSE2,	0,	0
 mulss,	xmm_xmm32_insn,	4,	SUF_Z,	0xF3,	0x59,	0,	0,	CPU_SSE,	0,	0
+mulx,	vex_gpr_reg_nds_rm_0F_insn,	2,	SUF_Z,	0xF2,	0x38,	0xF6,	ONLY_AVX,	CPU_BMI2,	0,	0
 mwait,	threebyte_insn,	1,	SUF_Z,	0x0F,	0x01,	0xC9,	0,	CPU_SSE3,	0,	0
 neg,	f6_insn,	4,	SUF_Z,	0x03,	0,	0,	0,	0,	0,	0
 nop,	onebyte_insn,	1,	SUF_Z,	0x90,	0,	0,	0,	0,	0,	0
@@ -442,9 +450,9 @@
 outsb,	onebyte_insn,	1,	SUF_Z,	0x6E,	0x00,	0,	0,	0,	0,	0
 outsd,	onebyte_insn,	1,	SUF_Z,	0x6F,	0x20,	0,	0,	CPU_386,	0,	0
 outsw,	onebyte_insn,	1,	SUF_Z,	0x6F,	0x10,	0,	0,	0,	0,	0
-pabsb,	ssse3_insn,	3,	SUF_Z,	0x1C,	0,	0,	0,	CPU_SSSE3,	0,	0
-pabsd,	ssse3_insn,	3,	SUF_Z,	0x1E,	0,	0,	0,	CPU_SSSE3,	0,	0
-pabsw,	ssse3_insn,	3,	SUF_Z,	0x1D,	0,	0,	0,	CPU_SSSE3,	0,	0
+pabsb,	ssse3_insn,	5,	SUF_Z,	0x1C,	0,	0,	0,	CPU_SSSE3,	0,	0
+pabsd,	ssse3_insn,	5,	SUF_Z,	0x1E,	0,	0,	0,	CPU_SSSE3,	0,	0
+pabsw,	ssse3_insn,	5,	SUF_Z,	0x1D,	0,	0,	0,	CPU_SSSE3,	0,	0
 packssdw,	mmxsse2_insn,	2,	SUF_Z,	0x6B,	0,	0,	0,	CPU_MMX,	0,	0
 packsswb,	mmxsse2_insn,	2,	SUF_Z,	0x63,	0,	0,	0,	CPU_MMX,	0,	0
 packusdw,	sse4_insn,	2,	SUF_Z,	0x2B,	0,	0,	0,	CPU_SSE41,	0,	0
@@ -485,7 +493,9 @@
 pcmpgtw,	mmxsse2_insn,	2,	SUF_Z,	0x65,	0,	0,	0,	CPU_MMX,	0,	0
 pcmpistri,	sse4pcmpstr_insn,	1,	SUF_Z,	0x63,	0,	0,	0,	CPU_SSE42,	0,	0
 pcmpistrm,	sse4pcmpstr_insn,	1,	SUF_Z,	0x62,	0,	0,	0,	CPU_SSE42,	0,	0
+pdep,	vex_gpr_reg_nds_rm_0F_insn,	2,	SUF_Z,	0xF2,	0x38,	0xF5,	ONLY_AVX,	CPU_BMI2,	0,	0
 pdistib,	cyrixmmx_insn,	1,	SUF_Z,	0x54,	0,	0,	0,	CPU_Cyrix,	CPU_MMX,	0
+pext,	vex_gpr_reg_nds_rm_0F_insn,	2,	SUF_Z,	0xF3,	0x38,	0xF5,	ONLY_AVX,	CPU_BMI2,	0,	0
 pextrb,	pextrb_insn,	3,	SUF_Z,	0,	0,	0,	0,	CPU_SSE41,	0,	0
 pextrd,	pextrd_insn,	1,	SUF_Z,	0,	0,	0,	0,	CPU_386,	CPU_SSE41,	0
 pextrq,	pextrq_insn,	1,	SUF_Z,	0,	0,	0,	ONLY_64,	CPU_SSE41,	0,	0
@@ -509,13 +519,13 @@
 pfrsqrt,	now3d_insn,	1,	SUF_Z,	0x97,	0,	0,	0,	CPU_3DNow,	0,	0
 pfsub,	now3d_insn,	1,	SUF_Z,	0x9A,	0,	0,	0,	CPU_3DNow,	0,	0
 pfsubr,	now3d_insn,	1,	SUF_Z,	0xAA,	0,	0,	0,	CPU_3DNow,	0,	0
-phaddd,	ssse3_insn,	3,	SUF_Z,	0x02,	0,	0,	0,	CPU_SSSE3,	0,	0
-phaddsw,	ssse3_insn,	3,	SUF_Z,	0x03,	0,	0,	0,	CPU_SSSE3,	0,	0
-phaddw,	ssse3_insn,	3,	SUF_Z,	0x01,	0,	0,	0,	CPU_SSSE3,	0,	0
+phaddd,	ssse3_insn,	5,	SUF_Z,	0x02,	0,	0,	0,	CPU_SSSE3,	0,	0
+phaddsw,	ssse3_insn,	5,	SUF_Z,	0x03,	0,	0,	0,	CPU_SSSE3,	0,	0
+phaddw,	ssse3_insn,	5,	SUF_Z,	0x01,	0,	0,	0,	CPU_SSSE3,	0,	0
 phminposuw,	sse4_insn,	2,	SUF_Z,	0x41,	0,	0,	0,	CPU_SSE41,	0,	0
-phsubd,	ssse3_insn,	3,	SUF_Z,	0x06,	0,	0,	0,	CPU_SSSE3,	0,	0
-phsubsw,	ssse3_insn,	3,	SUF_Z,	0x07,	0,	0,	0,	CPU_SSSE3,	0,	0
-phsubw,	ssse3_insn,	3,	SUF_Z,	0x05,	0,	0,	0,	CPU_SSSE3,	0,	0
+phsubd,	ssse3_insn,	5,	SUF_Z,	0x06,	0,	0,	0,	CPU_SSSE3,	0,	0
+phsubsw,	ssse3_insn,	5,	SUF_Z,	0x07,	0,	0,	0,	CPU_SSSE3,	0,	0
+phsubw,	ssse3_insn,	5,	SUF_Z,	0x05,	0,	0,	0,	CPU_SSSE3,	0,	0
 pi2fd,	now3d_insn,	1,	SUF_Z,	0x0D,	0,	0,	0,	CPU_3DNow,	0,	0
 pi2fw,	now3d_insn,	1,	SUF_Z,	0x0C,	0,	0,	0,	CPU_3DNow,	CPU_Athlon,	0
 pinsrb,	pinsrb_insn,	4,	SUF_Z,	0,	0,	0,	0,	CPU_SSE41,	0,	0
@@ -523,7 +533,7 @@
 pinsrq,	pinsrq_insn,	2,	SUF_Z,	0,	0,	0,	ONLY_64,	CPU_SSE41,	0,	0
 pinsrw,	pinsrw_insn,	9,	SUF_Z,	0,	0,	0,	0,	CPU_MMX,	CPU_P3,	0
 pmachriw,	pmachriw_insn,	1,	SUF_Z,	0,	0,	0,	0,	CPU_Cyrix,	CPU_MMX,	0
-pmaddubsw,	ssse3_insn,	3,	SUF_Z,	0x04,	0,	0,	0,	CPU_SSSE3,	0,	0
+pmaddubsw,	ssse3_insn,	5,	SUF_Z,	0x04,	0,	0,	0,	CPU_SSSE3,	0,	0
 pmaddwd,	mmxsse2_insn,	2,	SUF_Z,	0xF5,	0,	0,	0,	CPU_MMX,	0,	0
 pmagw,	cyrixmmx_insn,	1,	SUF_Z,	0x52,	0,	0,	0,	CPU_Cyrix,	CPU_MMX,	0
 pmaxsb,	sse4_insn,	2,	SUF_Z,	0x3C,	0,	0,	0,	CPU_SSE41,	0,	0
@@ -538,22 +548,22 @@
 pminub,	mmxsse2_insn,	2,	SUF_Z,	0xDA,	0,	0,	0,	CPU_MMX,	CPU_P3,	0
 pminud,	sse4_insn,	2,	SUF_Z,	0x3B,	0,	0,	0,	CPU_SSE41,	0,	0
 pminuw,	sse4_insn,	2,	SUF_Z,	0x3A,	0,	0,	0,	CPU_SSE41,	0,	0
-pmovmskb,	pmovmskb_insn,	4,	SUF_Z,	0,	0,	0,	0,	CPU_MMX,	CPU_P3,	0
-pmovsxbd,	sse4m32_insn,	2,	SUF_Z,	0x21,	0,	0,	0,	CPU_SSE41,	0,	0
-pmovsxbq,	sse4m16_insn,	2,	SUF_Z,	0x22,	0,	0,	0,	CPU_SSE41,	0,	0
-pmovsxbw,	sse4m64_insn,	2,	SUF_Z,	0x20,	0,	0,	0,	CPU_SSE41,	0,	0
-pmovsxdq,	sse4m64_insn,	2,	SUF_Z,	0x25,	0,	0,	0,	CPU_SSE41,	0,	0
-pmovsxwd,	sse4m64_insn,	2,	SUF_Z,	0x23,	0,	0,	0,	CPU_SSE41,	0,	0
-pmovsxwq,	sse4m32_insn,	2,	SUF_Z,	0x24,	0,	0,	0,	CPU_SSE41,	0,	0
-pmovzxbd,	sse4m32_insn,	2,	SUF_Z,	0x31,	0,	0,	0,	CPU_SSE41,	0,	0
-pmovzxbq,	sse4m16_insn,	2,	SUF_Z,	0x32,	0,	0,	0,	CPU_SSE41,	0,	0
-pmovzxbw,	sse4m64_insn,	2,	SUF_Z,	0x30,	0,	0,	0,	CPU_SSE41,	0,	0
-pmovzxdq,	sse4m64_insn,	2,	SUF_Z,	0x35,	0,	0,	0,	CPU_SSE41,	0,	0
-pmovzxwd,	sse4m64_insn,	2,	SUF_Z,	0x33,	0,	0,	0,	CPU_SSE41,	0,	0
-pmovzxwq,	sse4m32_insn,	2,	SUF_Z,	0x34,	0,	0,	0,	CPU_SSE41,	0,	0
+pmovmskb,	pmovmskb_insn,	6,	SUF_Z,	0,	0,	0,	0,	CPU_MMX,	CPU_P3,	0
+pmovsxbd,	sse4m32_insn,	4,	SUF_Z,	0x21,	0,	0,	0,	CPU_SSE41,	0,	0
+pmovsxbq,	sse4m16_insn,	4,	SUF_Z,	0x22,	0,	0,	0,	CPU_SSE41,	0,	0
+pmovsxbw,	sse4m64_insn,	4,	SUF_Z,	0x20,	0,	0,	0,	CPU_SSE41,	0,	0
+pmovsxdq,	sse4m64_insn,	4,	SUF_Z,	0x25,	0,	0,	0,	CPU_SSE41,	0,	0
+pmovsxwd,	sse4m64_insn,	4,	SUF_Z,	0x23,	0,	0,	0,	CPU_SSE41,	0,	0
+pmovsxwq,	sse4m32_insn,	4,	SUF_Z,	0x24,	0,	0,	0,	CPU_SSE41,	0,	0
+pmovzxbd,	sse4m32_insn,	4,	SUF_Z,	0x31,	0,	0,	0,	CPU_SSE41,	0,	0
+pmovzxbq,	sse4m16_insn,	4,	SUF_Z,	0x32,	0,	0,	0,	CPU_SSE41,	0,	0
+pmovzxbw,	sse4m64_insn,	4,	SUF_Z,	0x30,	0,	0,	0,	CPU_SSE41,	0,	0
+pmovzxdq,	sse4m64_insn,	4,	SUF_Z,	0x35,	0,	0,	0,	CPU_SSE41,	0,	0
+pmovzxwd,	sse4m64_insn,	4,	SUF_Z,	0x33,	0,	0,	0,	CPU_SSE41,	0,	0
+pmovzxwq,	sse4m32_insn,	4,	SUF_Z,	0x34,	0,	0,	0,	CPU_SSE41,	0,	0
 pmuldq,	sse4_insn,	2,	SUF_Z,	0x28,	0,	0,	0,	CPU_SSE41,	0,	0
 pmulhriw,	cyrixmmx_insn,	1,	SUF_Z,	0x5D,	0,	0,	0,	CPU_Cyrix,	CPU_MMX,	0
-pmulhrsw,	ssse3_insn,	3,	SUF_Z,	0x0B,	0,	0,	0,	CPU_SSSE3,	0,	0
+pmulhrsw,	ssse3_insn,	5,	SUF_Z,	0x0B,	0,	0,	0,	CPU_SSSE3,	0,	0
 pmulhrwa,	now3d_insn,	1,	SUF_Z,	0xB7,	0,	0,	0,	CPU_3DNow,	0,	0
 pmulhrwc,	cyrixmmx_insn,	1,	SUF_Z,	0x59,	0,	0,	0,	CPU_Cyrix,	CPU_MMX,	0
 pmulhuw,	mmxsse2_insn,	2,	SUF_Z,	0xE4,	0,	0,	0,	CPU_MMX,	CPU_P3,	0
@@ -565,7 +575,7 @@
 pmvlzb,	cyrixmmx_insn,	1,	SUF_Z,	0x5B,	0,	0,	0,	CPU_Cyrix,	CPU_MMX,	0
 pmvnzb,	cyrixmmx_insn,	1,	SUF_Z,	0x5A,	0,	0,	0,	CPU_Cyrix,	CPU_MMX,	0
 pmvzb,	cyrixmmx_insn,	1,	SUF_Z,	0x58,	0,	0,	0,	CPU_Cyrix,	CPU_MMX,	0
-pop,	pop_insn,	21,	SUF_Z,	0,	0,	0,	0,	0,	0,	0
+pop,	pop_insn,	23,	SUF_Z,	0,	0,	0,	0,	0,	0,	0
 popa,	onebyte_insn,	1,	SUF_Z,	0x61,	0x00,	0,	NOT_64,	CPU_186,	0,	0
 popad,	onebyte_insn,	1,	SUF_Z,	0x61,	0x20,	0,	NOT_64,	CPU_386,	0,	0
 popaw,	onebyte_insn,	1,	SUF_Z,	0x61,	0x10,	0,	NOT_64,	CPU_186,	0,	0
@@ -582,22 +592,22 @@
 prefetcht2,	twobytemem_insn,	1,	SUF_Z,	0x03,	0x0F,	0x18,	0,	CPU_P3,	0,	0
 prefetchw,	twobytemem_insn,	1,	SUF_Z,	0x01,	0x0F,	0x0D,	0,	CPU_3DNow,	0,	0
 psadbw,	mmxsse2_insn,	2,	SUF_Z,	0xF6,	0,	0,	0,	CPU_MMX,	CPU_P3,	0
-pshufb,	ssse3_insn,	3,	SUF_Z,	0x00,	0,	0,	0,	CPU_SSSE3,	0,	0
+pshufb,	ssse3_insn,	5,	SUF_Z,	0x00,	0,	0,	0,	CPU_SSSE3,	0,	0
 pshufd,	xmm_xmm128_imm_insn,	1,	SUF_Z,	0x66,	0x70,	0,	0,	CPU_SSE2,	0,	0
 pshufhw,	xmm_xmm128_imm_insn,	1,	SUF_Z,	0xF3,	0x70,	0,	0,	CPU_SSE2,	0,	0
 pshuflw,	xmm_xmm128_imm_insn,	1,	SUF_Z,	0xF2,	0x70,	0,	0,	CPU_SSE2,	0,	0
 pshufw,	pshufw_insn,	1,	SUF_Z,	0,	0,	0,	0,	CPU_MMX,	CPU_P3,	0
-psignb,	ssse3_insn,	3,	SUF_Z,	0x08,	0,	0,	0,	CPU_SSSE3,	0,	0
-psignd,	ssse3_insn,	3,	SUF_Z,	0x0A,	0,	0,	0,	CPU_SSSE3,	0,	0
-psignw,	ssse3_insn,	3,	SUF_Z,	0x09,	0,	0,	0,	CPU_SSSE3,	0,	0
+psignb,	ssse3_insn,	5,	SUF_Z,	0x08,	0,	0,	0,	CPU_SSSE3,	0,	0
+psignd,	ssse3_insn,	5,	SUF_Z,	0x0A,	0,	0,	0,	CPU_SSSE3,	0,	0
+psignw,	ssse3_insn,	5,	SUF_Z,	0x09,	0,	0,	0,	CPU_SSSE3,	0,	0
 pslld,	pshift_insn,	4,	SUF_Z,	0xF2,	0x72,	0x06,	0,	CPU_MMX,	0,	0
-pslldq,	pslrldq_insn,	2,	SUF_Z,	0x07,	0,	0,	0,	CPU_SSE2,	0,	0
+pslldq,	pslrldq_insn,	4,	SUF_Z,	0x07,	0,	0,	0,	CPU_SSE2,	0,	0
 psllq,	pshift_insn,	4,	SUF_Z,	0xF3,	0x73,	0x06,	0,	CPU_MMX,	0,	0
 psllw,	pshift_insn,	4,	SUF_Z,	0xF1,	0x71,	0x06,	0,	CPU_MMX,	0,	0
 psrad,	pshift_insn,	4,	SUF_Z,	0xE2,	0x72,	0x04,	0,	CPU_MMX,	0,	0
 psraw,	pshift_insn,	4,	SUF_Z,	0xE1,	0x71,	0x04,	0,	CPU_MMX,	0,	0
 psrld,	pshift_insn,	4,	SUF_Z,	0xD2,	0x72,	0x02,	0,	CPU_MMX,	0,	0
-psrldq,	pslrldq_insn,	2,	SUF_Z,	0x03,	0,	0,	0,	CPU_SSE2,	0,	0
+psrldq,	pslrldq_insn,	4,	SUF_Z,	0x03,	0,	0,	0,	CPU_SSE2,	0,	0
 psrlq,	pshift_insn,	4,	SUF_Z,	0xD3,	0x73,	0x02,	0,	CPU_MMX,	0,	0
 psrlw,	pshift_insn,	4,	SUF_Z,	0xD1,	0x71,	0x02,	0,	CPU_MMX,	0,	0
 psubb,	mmxsse2_insn,	2,	SUF_Z,	0xF8,	0,	0,	0,	CPU_MMX,	0,	0
@@ -619,7 +629,7 @@
 punpckldq,	mmxsse2_insn,	2,	SUF_Z,	0x62,	0,	0,	0,	CPU_MMX,	0,	0
 punpcklqdq,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x6C,	0,	0,	CPU_SSE2,	0,	0
 punpcklwd,	mmxsse2_insn,	2,	SUF_Z,	0x61,	0,	0,	0,	CPU_MMX,	0,	0
-push,	push_insn,	33,	SUF_Z,	0,	0,	0,	0,	0,	0,	0
+push,	push_insn,	35,	SUF_Z,	0,	0,	0,	0,	0,	0,	0
 pusha,	onebyte_insn,	1,	SUF_Z,	0x60,	0x00,	0,	NOT_64,	CPU_186,	0,	0
 pushad,	onebyte_insn,	1,	SUF_Z,	0x60,	0x20,	0,	NOT_64,	CPU_386,	0,	0
 pushaw,	onebyte_insn,	1,	SUF_Z,	0x60,	0x10,	0,	NOT_64,	CPU_186,	0,	0
@@ -650,6 +660,7 @@
 retn,	retnf_insn,	6,	SUF_Z,	0xC2,	0,	0,	0,	0,	0,	0
 rol,	shift_insn,	16,	SUF_Z,	0x00,	0,	0,	0,	0,	0,	0
 ror,	shift_insn,	16,	SUF_Z,	0x01,	0,	0,	0,	0,	0,	0
+rorx,	vex_gpr_reg_rm_0F_imm8_insn,	2,	SUF_Z,	0xF2,	0x3A,	0xF0,	ONLY_AVX,	CPU_BMI2,	0,	0
 roundpd,	sse4imm_insn,	2,	SUF_Z,	0x09,	0,	0,	0,	CPU_SSE41,	0,	0
 roundps,	sse4imm_insn,	2,	SUF_Z,	0x08,	0,	0,	0,	CPU_SSE41,	0,	0
 roundsd,	sse4m64imm_insn,	4,	SUF_Z,	0x0B,	0,	0,	0,	CPU_SSE41,	0,	0
@@ -664,6 +675,7 @@
 sal,	shift_insn,	16,	SUF_Z,	0x04,	0,	0,	0,	0,	0,	0
 salc,	onebyte_insn,	1,	SUF_Z,	0xD6,	0,	0,	NOT_64,	CPU_Undoc,	0,	0
 sar,	shift_insn,	16,	SUF_Z,	0x07,	0,	0,	0,	0,	0,	0
+sarx,	vex_gpr_reg_rm_nds_0F_insn,	2,	SUF_Z,	0xF3,	0x38,	0xF7,	ONLY_AVX,	CPU_BMI2,	0,	0
 sbb,	arith_insn,	22,	SUF_Z,	0x18,	0x03,	0,	0,	0,	0,	0
 scasb,	onebyte_insn,	1,	SUF_Z,	0xAE,	0x00,	0,	0,	0,	0,	0
 scasd,	onebyte_insn,	1,	SUF_Z,	0xAF,	0x20,	0,	0,	CPU_386,	0,	0
@@ -703,8 +715,10 @@
 sgdt,	twobytemem_insn,	1,	SUF_Z,	0x00,	0x0F,	0x01,	0,	CPU_286,	CPU_Priv,	0
 shl,	shift_insn,	16,	SUF_Z,	0x04,	0,	0,	0,	0,	0,	0
 shld,	shlrd_insn,	9,	SUF_Z,	0xA4,	0,	0,	0,	CPU_386,	0,	0
+shlx,	vex_gpr_reg_rm_nds_0F_insn,	2,	SUF_Z,	0x66,	0x38,	0xF7,	ONLY_AVX,	CPU_BMI2,	0,	0
 shr,	shift_insn,	16,	SUF_Z,	0x05,	0,	0,	0,	0,	0,	0
 shrd,	shlrd_insn,	9,	SUF_Z,	0xAC,	0,	0,	0,	CPU_386,	0,	0
+shrx,	vex_gpr_reg_rm_nds_0F_insn,	2,	SUF_Z,	0xF2,	0x38,	0xF7,	ONLY_AVX,	CPU_BMI2,	0,	0
 shufpd,	xmm_xmm128_imm_insn,	1,	SUF_Z,	0x66,	0xC6,	0,	0,	CPU_SSE2,	0,	0
 shufps,	xmm_xmm128_imm_insn,	1,	SUF_Z,	0x00,	0xC6,	0,	0,	CPU_SSE,	0,	0
 sidt,	twobytemem_insn,	1,	SUF_Z,	0x01,	0x0F,	0x01,	0,	CPU_286,	CPU_Priv,	0
@@ -742,6 +756,7 @@
 sysexit,	twobyte_insn,	1,	SUF_Z,	0x0F,	0x35,	0,	NOT_64,	CPU_686,	CPU_Priv,	0
 sysret,	twobyte_insn,	1,	SUF_Z,	0x0F,	0x07,	0,	0,	CPU_686,	CPU_AMD,	CPU_Priv
 test,	test_insn,	20,	SUF_Z,	0,	0,	0,	0,	0,	0,	0
+tzcnt,	cnt_insn,	3,	SUF_Z,	0xBC,	0,	0,	0,	CPU_BMI1,	0,	0
 ucomisd,	xmm_xmm64_insn,	4,	SUF_Z,	0x66,	0x2E,	0,	0,	CPU_SSE2,	0,	0
 ucomiss,	xmm_xmm32_insn,	4,	SUF_Z,	0x00,	0x2E,	0,	0,	CPU_SSE,	0,	0
 ud1,	twobyte_insn,	1,	SUF_Z,	0x0F,	0xB9,	0,	0,	CPU_286,	CPU_Undoc,	0
@@ -751,29 +766,30 @@
 unpckhps,	xmm_xmm128_insn,	2,	SUF_Z,	0x00,	0x15,	0,	0,	CPU_SSE,	0,	0
 unpcklpd,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x14,	0,	0,	CPU_SSE2,	0,	0
 unpcklps,	xmm_xmm128_insn,	2,	SUF_Z,	0x00,	0x14,	0,	0,	CPU_SSE,	0,	0
-vaddpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x58,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vaddps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x00,	0x58,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vaddpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x58,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vaddps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x00,	0x58,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vaddsd,	xmm_xmm64_insn,	4,	SUF_Z,	0xF2,	0x58,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vaddss,	xmm_xmm32_insn,	4,	SUF_Z,	0xF3,	0x58,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vaddsubpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0xD0,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vaddsubps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0xF2,	0xD0,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vaddsubpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0xD0,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vaddsubps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0xF2,	0xD0,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vaesdec,	aes_insn,	2,	SUF_Z,	0x38,	0xDE,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vaesdeclast,	aes_insn,	2,	SUF_Z,	0x38,	0xDF,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vaesenc,	aes_insn,	2,	SUF_Z,	0x38,	0xDC,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vaesenclast,	aes_insn,	2,	SUF_Z,	0x38,	0xDD,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vaesimc,	aesimc_insn,	1,	SUF_Z,	0x38,	0xDB,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vaeskeygenassist,	aes_imm_insn,	1,	SUF_Z,	0x3A,	0xDF,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vandnpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x55,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vandnps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x00,	0x55,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vandpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x54,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vandps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x00,	0x54,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vblendpd,	sse4imm_256_insn,	3,	SUF_Z,	0x0D,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vblendps,	sse4imm_256_insn,	3,	SUF_Z,	0x0C,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vandnpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x55,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vandnps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x00,	0x55,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vandpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x54,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vandps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x00,	0x54,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vblendpd,	sse4imm_256_insn,	4,	SUF_Z,	0x0D,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vblendps,	sse4imm_256_insn,	4,	SUF_Z,	0x0C,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vblendvpd,	avx_sse4xmm0_insn,	2,	SUF_Z,	0x4B,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vblendvps,	avx_sse4xmm0_insn,	2,	SUF_Z,	0x4A,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vbroadcastf128,	vbroadcastf128_insn,	1,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vbroadcastsd,	vbroadcastsd_insn,	1,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vbroadcastss,	vbroadcastss_insn,	2,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vbroadcastf128,	vbroadcastif128_insn,	1,	SUF_Z,	0x1A,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vbroadcasti128,	vbroadcastif128_insn,	1,	SUF_Z,	0x5A,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vbroadcastsd,	vbroadcastsd_insn,	2,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vbroadcastss,	vbroadcastss_insn,	4,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vcmpeq_ospd,	ssecmp_128_insn,	3,	SUF_Z,	0x10,	0x66,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vcmpeq_osps,	ssecmp_128_insn,	3,	SUF_Z,	0x10,	0x00,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vcmpeq_ossd,	ssecmp_64_insn,	4,	SUF_Z,	0x10,	0xF2,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
@@ -926,15 +942,16 @@
 vcvttps2dq,	avx_xmm_xmm128_insn,	2,	SUF_Z,	0xF3,	0x5B,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vcvttsd2si,	cvt_rx_xmm64_insn,	4,	SUF_Z,	0xF2,	0x2C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vcvttss2si,	cvt_rx_xmm32_insn,	4,	SUF_Z,	0xF3,	0x2C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vdivpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x5E,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vdivps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x00,	0x5E,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vdivpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x5E,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vdivps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x00,	0x5E,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vdivsd,	xmm_xmm64_insn,	4,	SUF_Z,	0xF2,	0x5E,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vdivss,	xmm_xmm32_insn,	4,	SUF_Z,	0xF3,	0x5E,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vdppd,	sse4imm_insn,	2,	SUF_Z,	0x41,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vdpps,	sse4imm_256_insn,	3,	SUF_Z,	0x40,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vdpps,	sse4imm_256_insn,	4,	SUF_Z,	0x40,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 verr,	prot286_insn,	1,	SUF_Z,	0x04,	0x00,	0,	0,	CPU_286,	CPU_Prot,	0
 verw,	prot286_insn,	1,	SUF_Z,	0x05,	0x00,	0,	0,	CPU_286,	CPU_Prot,	0
-vextractf128,	vextractf128_insn,	1,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vextractf128,	vextractif128_insn,	1,	SUF_Z,	0x19,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vextracti128,	vextractif128_insn,	1,	SUF_Z,	0x39,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
 vextractps,	extractps_insn,	2,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vfmadd132pd,	vfma_pd_insn,	2,	SUF_Z,	0x98,	0,	0,	ONLY_AVX,	CPU_FMA,	0,	0
 vfmadd132ps,	vfma_ps_insn,	2,	SUF_Z,	0x98,	0,	0,	ONLY_AVX,	CPU_FMA,	0,	0
@@ -1020,25 +1037,30 @@
 vfrczps,	vfrc_pdps_insn,	2,	SUF_Z,	0x00,	0,	0,	0,	CPU_XOP,	0,	0
 vfrczsd,	vfrczsd_insn,	2,	SUF_Z,	0,	0,	0,	0,	CPU_XOP,	0,	0
 vfrczss,	vfrczss_insn,	2,	SUF_Z,	0,	0,	0,	0,	CPU_XOP,	0,	0
-vhaddpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x7C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vhaddps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0xF2,	0x7C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vhsubpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x7D,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vhsubps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0xF2,	0x7D,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vinsertf128,	vinsertf128_insn,	1,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vgatherdpd,	gather_64x_64x_insn,	2,	SUF_Z,	0x92,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vgatherdps,	gather_32x_32y_insn,	2,	SUF_Z,	0x92,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vgatherqpd,	gather_64x_64y_insn,	2,	SUF_Z,	0x93,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vgatherqps,	gather_32x_32y_128_insn,	2,	SUF_Z,	0x93,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vhaddpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x7C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vhaddps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0xF2,	0x7C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vhsubpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x7D,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vhsubps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0xF2,	0x7D,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vinsertf128,	vinsertif128_insn,	1,	SUF_Z,	0x18,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vinserti128,	vinsertif128_insn,	1,	SUF_Z,	0x38,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
 vinsertps,	insertps_insn,	4,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vlddqu,	lddqu_insn,	2,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vldmxcsr,	ldstmxcsr_insn,	1,	SUF_Z,	0x02,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmaskmovdqu,	maskmovdqu_insn,	1,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmaskmovpd,	vmaskmov_insn,	4,	SUF_Z,	0x2D,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmaskmovps,	vmaskmov_insn,	4,	SUF_Z,	0x2C,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vmaxpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x5F,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vmaxps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x00,	0x5F,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vmaxpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x5F,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vmaxps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x00,	0x5F,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmaxsd,	xmm_xmm64_insn,	4,	SUF_Z,	0xF2,	0x5F,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmaxss,	xmm_xmm32_insn,	4,	SUF_Z,	0xF3,	0x5F,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmcall,	threebyte_insn,	1,	SUF_Z,	0x0F,	0x01,	0xC1,	0,	CPU_P4,	0,	0
 vmclear,	vmxthreebytemem_insn,	1,	SUF_Z,	0x66,	0,	0,	0,	CPU_P4,	0,	0
-vminpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x5D,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vminps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x00,	0x5D,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vminpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x5D,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vminps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x00,	0x5D,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vminsd,	xmm_xmm64_insn,	4,	SUF_Z,	0xF2,	0x5D,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vminss,	xmm_xmm32_insn,	4,	SUF_Z,	0xF3,	0x5D,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmlaunch,	threebyte_insn,	1,	SUF_Z,	0x0F,	0x01,	0xC2,	0,	CPU_P4,	0,	0
@@ -1059,7 +1081,7 @@
 vmovmskpd,	movmsk_insn,	4,	SUF_Z,	0x66,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmovmskps,	movmsk_insn,	4,	SUF_Z,	0x00,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmovntdq,	movnt_insn,	2,	SUF_Z,	0x66,	0xE7,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vmovntdqa,	movntdqa_insn,	1,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vmovntdqa,	movntdqa_insn,	2,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmovntpd,	movnt_insn,	2,	SUF_Z,	0x66,	0x2B,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmovntps,	movnt_insn,	2,	SUF_Z,	0x00,	0x2B,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmovq,	vmovq_insn,	5,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
@@ -1069,60 +1091,65 @@
 vmovss,	movss_insn,	4,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmovupd,	movau_insn,	6,	SUF_Z,	0x66,	0x10,	0x01,	ONLY_AVX,	CPU_AVX,	0,	0
 vmovups,	movau_insn,	6,	SUF_Z,	0x00,	0x10,	0x01,	ONLY_AVX,	CPU_AVX,	0,	0
-vmpsadbw,	sse4imm_insn,	2,	SUF_Z,	0x42,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vmpsadbw,	sse4imm_256avx2_insn,	4,	SUF_Z,	0x42,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmptrld,	vmxtwobytemem_insn,	1,	SUF_Z,	0x06,	0,	0,	0,	CPU_P4,	0,	0
 vmptrst,	vmxtwobytemem_insn,	1,	SUF_Z,	0x07,	0,	0,	0,	CPU_P4,	0,	0
 vmread,	vmxmemrd_insn,	2,	SUF_Z,	0,	0,	0,	0,	CPU_P4,	0,	0
 vmresume,	threebyte_insn,	1,	SUF_Z,	0x0F,	0x01,	0xC3,	0,	CPU_P4,	0,	0
 vmrun,	svm_rax_insn,	2,	SUF_Z,	0xD8,	0,	0,	0,	CPU_SVM,	0,	0
 vmsave,	svm_rax_insn,	2,	SUF_Z,	0xDB,	0,	0,	0,	CPU_SVM,	0,	0
-vmulpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x59,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vmulps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x00,	0x59,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vmulpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x59,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vmulps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x00,	0x59,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmulsd,	xmm_xmm64_insn,	4,	SUF_Z,	0xF2,	0x59,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmulss,	xmm_xmm32_insn,	4,	SUF_Z,	0xF3,	0x59,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vmwrite,	vmxmemwr_insn,	2,	SUF_Z,	0,	0,	0,	0,	CPU_P4,	0,	0
 vmxoff,	threebyte_insn,	1,	SUF_Z,	0x0F,	0x01,	0xC4,	0,	CPU_P4,	0,	0
 vmxon,	vmxthreebytemem_insn,	1,	SUF_Z,	0xF3,	0,	0,	0,	CPU_P4,	0,	0
-vorpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x56,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vorps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x00,	0x56,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpabsb,	avx_ssse3_2op_insn,	1,	SUF_Z,	0x1C,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpabsd,	avx_ssse3_2op_insn,	1,	SUF_Z,	0x1E,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpabsw,	avx_ssse3_2op_insn,	1,	SUF_Z,	0x1D,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpackssdw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x6B,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpacksswb,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x63,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpackusdw,	ssse3_insn,	3,	SUF_Z,	0x2B,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpackuswb,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x67,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpaddb,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xFC,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpaddd,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xFE,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpaddq,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xD4,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpaddsb,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xEC,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpaddsw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xED,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpaddusb,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xDC,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpaddusw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xDD,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpaddw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xFD,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpalignr,	sse4imm_insn,	2,	SUF_Z,	0x0F,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpand,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xDB,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpandn,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xDF,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpavgb,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xE0,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpavgw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xE3,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpblendvb,	avx_sse4xmm0_128_insn,	1,	SUF_Z,	0x4C,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpblendw,	sse4imm_insn,	2,	SUF_Z,	0x0E,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vorpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x56,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vorps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x00,	0x56,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpabsb,	avx2_ssse3_2op_insn,	2,	SUF_Z,	0x1C,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpabsd,	avx2_ssse3_2op_insn,	2,	SUF_Z,	0x1E,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpabsw,	avx2_ssse3_2op_insn,	2,	SUF_Z,	0x1D,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpackssdw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x6B,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpacksswb,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x63,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpackusdw,	ssse3_insn,	5,	SUF_Z,	0x2B,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpackuswb,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x67,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpaddb,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xFC,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpaddd,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xFE,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpaddq,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xD4,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpaddsb,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xEC,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpaddsw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xED,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpaddusb,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xDC,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpaddusw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xDD,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpaddw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xFD,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpalignr,	sse4imm_256avx2_insn,	4,	SUF_Z,	0x0F,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpand,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xDB,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpandn,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xDF,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpavgb,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xE0,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpavgw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xE3,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpblendd,	vex_66_0F3A_imm8_avx2_insn,	2,	SUF_Z,	0x02,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpblendvb,	avx2_sse4xmm0_insn,	2,	SUF_Z,	0x4C,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpblendw,	sse4imm_256avx2_insn,	4,	SUF_Z,	0x0E,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpbroadcastb,	vpbroadcastb_avx2_insn,	4,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpbroadcastd,	vpbroadcastd_avx2_insn,	4,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpbroadcastq,	vpbroadcastq_avx2_insn,	4,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpbroadcastw,	vpbroadcastw_avx2_insn,	4,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
 vpclmulhqhqdq,	pclmulqdq_fixed_insn,	2,	SUF_Z,	0x11,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpclmulhqlqdq,	pclmulqdq_fixed_insn,	2,	SUF_Z,	0x01,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpclmullqhqdq,	pclmulqdq_fixed_insn,	2,	SUF_Z,	0x10,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpclmullqlqdq,	pclmulqdq_fixed_insn,	2,	SUF_Z,	0x00,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpclmulqdq,	pclmulqdq_insn,	2,	SUF_Z,	0x3A,	0x44,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpcmov,	vpcmov_insn,	4,	SUF_Z,	0,	0,	0,	0,	CPU_XOP,	0,	0
-vpcmpeqb,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x74,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpcmpeqd,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x76,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpcmpeqq,	ssse3_insn,	3,	SUF_Z,	0x29,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpcmpeqw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x75,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpcmpeqb,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x74,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpcmpeqd,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x76,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpcmpeqq,	ssse3_insn,	5,	SUF_Z,	0x29,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpcmpeqw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x75,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpcmpestri,	sse4pcmpstr_insn,	1,	SUF_Z,	0x61,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpcmpestrm,	sse4pcmpstr_insn,	1,	SUF_Z,	0x60,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpcmpgtb,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x64,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpcmpgtd,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x66,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpcmpgtq,	ssse3_insn,	3,	SUF_Z,	0x37,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpcmpgtw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x65,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpcmpgtb,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x64,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpcmpgtd,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x66,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpcmpgtq,	ssse3_insn,	5,	SUF_Z,	0x37,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpcmpgtw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x65,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpcmpistri,	sse4pcmpstr_insn,	1,	SUF_Z,	0x63,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpcmpistrm,	sse4pcmpstr_insn,	1,	SUF_Z,	0x62,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpcomb,	vpcom_imm_insn,	1,	SUF_Z,	0xCC,	0,	0,	0,	CPU_XOP,	0,	0
@@ -1206,33 +1233,42 @@
 vpcomuw,	vpcom_imm_insn,	1,	SUF_Z,	0xED,	0,	0,	0,	CPU_XOP,	0,	0
 vpcomw,	vpcom_imm_insn,	1,	SUF_Z,	0xCD,	0,	0,	0,	CPU_XOP,	0,	0
 vperm2f128,	vperm2f128_insn,	1,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vperm2i128,	vperm2i128_avx2_insn,	1,	SUF_Z,	0,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpermd,	vperm_var_avx2_insn,	1,	SUF_Z,	0x36,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
 vpermilpd,	vpermil_insn,	4,	SUF_Z,	0x05,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpermilps,	vpermil_insn,	4,	SUF_Z,	0x04,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpermpd,	vperm_imm_avx2_insn,	1,	SUF_Z,	0x01,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpermps,	vperm_var_avx2_insn,	1,	SUF_Z,	0x16,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpermq,	vperm_imm_avx2_insn,	1,	SUF_Z,	0x00,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
 vpextrb,	pextrb_insn,	3,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpextrd,	pextrd_insn,	1,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpextrq,	pextrq_insn,	1,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpextrw,	pextrw_insn,	7,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpgatherdd,	gather_32x_32y_insn,	2,	SUF_Z,	0x90,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpgatherdq,	gather_64x_64x_insn,	2,	SUF_Z,	0x90,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpgatherqd,	gather_32x_32y_128_insn,	2,	SUF_Z,	0x91,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpgatherqq,	gather_64x_64y_insn,	2,	SUF_Z,	0x91,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
 vphaddbd,	vphaddsub_insn,	1,	SUF_Z,	0xC2,	0,	0,	0,	CPU_XOP,	0,	0
 vphaddbq,	vphaddsub_insn,	1,	SUF_Z,	0xC3,	0,	0,	0,	CPU_XOP,	0,	0
 vphaddbw,	vphaddsub_insn,	1,	SUF_Z,	0xC1,	0,	0,	0,	CPU_XOP,	0,	0
-vphaddd,	ssse3_insn,	3,	SUF_Z,	0x02,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vphaddd,	ssse3_insn,	5,	SUF_Z,	0x02,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vphadddq,	vphaddsub_insn,	1,	SUF_Z,	0xCB,	0,	0,	0,	CPU_XOP,	0,	0
-vphaddsw,	ssse3_insn,	3,	SUF_Z,	0x03,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vphaddsw,	ssse3_insn,	5,	SUF_Z,	0x03,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vphaddubd,	vphaddsub_insn,	1,	SUF_Z,	0xD2,	0,	0,	0,	CPU_XOP,	0,	0
 vphaddubq,	vphaddsub_insn,	1,	SUF_Z,	0xD3,	0,	0,	0,	CPU_XOP,	0,	0
 vphaddubw,	vphaddsub_insn,	1,	SUF_Z,	0xD1,	0,	0,	0,	CPU_XOP,	0,	0
 vphaddudq,	vphaddsub_insn,	1,	SUF_Z,	0xD8,	0,	0,	0,	CPU_XOP,	0,	0
 vphadduwd,	vphaddsub_insn,	1,	SUF_Z,	0xD6,	0,	0,	0,	CPU_XOP,	0,	0
 vphadduwq,	vphaddsub_insn,	1,	SUF_Z,	0xD7,	0,	0,	0,	CPU_XOP,	0,	0
-vphaddw,	ssse3_insn,	3,	SUF_Z,	0x01,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vphaddw,	ssse3_insn,	5,	SUF_Z,	0x01,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vphaddwd,	vphaddsub_insn,	1,	SUF_Z,	0xC6,	0,	0,	0,	CPU_XOP,	0,	0
 vphaddwq,	vphaddsub_insn,	1,	SUF_Z,	0xC7,	0,	0,	0,	CPU_XOP,	0,	0
 vphminposuw,	avx_ssse3_2op_insn,	1,	SUF_Z,	0x41,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vphsubbw,	vphaddsub_insn,	1,	SUF_Z,	0xE1,	0,	0,	0,	CPU_XOP,	0,	0
-vphsubd,	ssse3_insn,	3,	SUF_Z,	0x06,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vphsubd,	ssse3_insn,	5,	SUF_Z,	0x06,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vphsubdq,	vphaddsub_insn,	1,	SUF_Z,	0xE3,	0,	0,	0,	CPU_XOP,	0,	0
-vphsubsw,	ssse3_insn,	3,	SUF_Z,	0x07,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vphsubw,	ssse3_insn,	3,	SUF_Z,	0x05,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vphsubsw,	ssse3_insn,	5,	SUF_Z,	0x07,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vphsubw,	ssse3_insn,	5,	SUF_Z,	0x05,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vphsubwd,	vphaddsub_insn,	1,	SUF_Z,	0xE2,	0,	0,	0,	CPU_XOP,	0,	0
 vpinsrb,	pinsrb_insn,	4,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpinsrd,	pinsrd_insn,	2,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
@@ -1250,47 +1286,49 @@
 vpmacsww,	vpma_insn,	1,	SUF_Z,	0x95,	0,	0,	0,	CPU_XOP,	0,	0
 vpmadcsswd,	vpma_insn,	1,	SUF_Z,	0xA6,	0,	0,	0,	CPU_XOP,	0,	0
 vpmadcswd,	vpma_insn,	1,	SUF_Z,	0xB6,	0,	0,	0,	CPU_XOP,	0,	0
-vpmaddubsw,	ssse3_insn,	3,	SUF_Z,	0x04,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmaddwd,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xF5,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmaxsb,	ssse3_insn,	3,	SUF_Z,	0x3C,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmaxsd,	ssse3_insn,	3,	SUF_Z,	0x3D,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmaxsw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xEE,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmaxub,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xDE,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmaxud,	ssse3_insn,	3,	SUF_Z,	0x3F,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmaxuw,	ssse3_insn,	3,	SUF_Z,	0x3E,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpminsb,	ssse3_insn,	3,	SUF_Z,	0x38,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpminsd,	ssse3_insn,	3,	SUF_Z,	0x39,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpminsw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xEA,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpminub,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xDA,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpminud,	ssse3_insn,	3,	SUF_Z,	0x3B,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpminuw,	ssse3_insn,	3,	SUF_Z,	0x3A,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovmskb,	pmovmskb_insn,	4,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovsxbd,	sse4m32_insn,	2,	SUF_Z,	0x21,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovsxbq,	sse4m16_insn,	2,	SUF_Z,	0x22,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovsxbw,	sse4m64_insn,	2,	SUF_Z,	0x20,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovsxdq,	sse4m64_insn,	2,	SUF_Z,	0x25,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovsxwd,	sse4m64_insn,	2,	SUF_Z,	0x23,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovsxwq,	sse4m32_insn,	2,	SUF_Z,	0x24,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovzxbd,	sse4m32_insn,	2,	SUF_Z,	0x31,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovzxbq,	sse4m16_insn,	2,	SUF_Z,	0x32,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovzxbw,	sse4m64_insn,	2,	SUF_Z,	0x30,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovzxdq,	sse4m64_insn,	2,	SUF_Z,	0x35,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovzxwd,	sse4m64_insn,	2,	SUF_Z,	0x33,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmovzxwq,	sse4m32_insn,	2,	SUF_Z,	0x34,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmuldq,	ssse3_insn,	3,	SUF_Z,	0x28,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmulhrsw,	ssse3_insn,	3,	SUF_Z,	0x0B,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmulhuw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xE4,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmulhw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xE5,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmulld,	ssse3_insn,	3,	SUF_Z,	0x40,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmullw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xD5,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpmuludq,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xF4,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpor,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xEB,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmaddubsw,	ssse3_insn,	5,	SUF_Z,	0x04,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmaddwd,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xF5,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmaskmovd,	vmaskmov_insn,	4,	SUF_Z,	0x8C,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpmaskmovq,	vmaskmov_vexw1_avx2_insn,	4,	SUF_Z,	0x8C,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpmaxsb,	ssse3_insn,	5,	SUF_Z,	0x3C,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmaxsd,	ssse3_insn,	5,	SUF_Z,	0x3D,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmaxsw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xEE,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmaxub,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xDE,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmaxud,	ssse3_insn,	5,	SUF_Z,	0x3F,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmaxuw,	ssse3_insn,	5,	SUF_Z,	0x3E,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpminsb,	ssse3_insn,	5,	SUF_Z,	0x38,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpminsd,	ssse3_insn,	5,	SUF_Z,	0x39,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpminsw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xEA,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpminub,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xDA,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpminud,	ssse3_insn,	5,	SUF_Z,	0x3B,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpminuw,	ssse3_insn,	5,	SUF_Z,	0x3A,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovmskb,	pmovmskb_insn,	6,	SUF_Z,	0xC0,	0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovsxbd,	sse4m32_insn,	4,	SUF_Z,	0x21,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovsxbq,	sse4m16_insn,	4,	SUF_Z,	0x22,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovsxbw,	sse4m64_insn,	4,	SUF_Z,	0x20,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovsxdq,	sse4m64_insn,	4,	SUF_Z,	0x25,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovsxwd,	sse4m64_insn,	4,	SUF_Z,	0x23,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovsxwq,	sse4m32_insn,	4,	SUF_Z,	0x24,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovzxbd,	sse4m32_insn,	4,	SUF_Z,	0x31,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovzxbq,	sse4m16_insn,	4,	SUF_Z,	0x32,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovzxbw,	sse4m64_insn,	4,	SUF_Z,	0x30,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovzxdq,	sse4m64_insn,	4,	SUF_Z,	0x35,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovzxwd,	sse4m64_insn,	4,	SUF_Z,	0x33,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmovzxwq,	sse4m32_insn,	4,	SUF_Z,	0x34,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmuldq,	ssse3_insn,	5,	SUF_Z,	0x28,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmulhrsw,	ssse3_insn,	5,	SUF_Z,	0x0B,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmulhuw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xE4,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmulhw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xE5,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmulld,	ssse3_insn,	5,	SUF_Z,	0x40,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmullw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xD5,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpmuludq,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xF4,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpor,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xEB,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpperm,	vpperm_insn,	2,	SUF_Z,	0,	0,	0,	0,	CPU_XOP,	0,	0
 vprotb,	vprot_insn,	3,	SUF_Z,	0x00,	0,	0,	0,	CPU_XOP,	0,	0
 vprotd,	vprot_insn,	3,	SUF_Z,	0x02,	0,	0,	0,	CPU_XOP,	0,	0
 vprotq,	vprot_insn,	3,	SUF_Z,	0x03,	0,	0,	0,	CPU_XOP,	0,	0
 vprotw,	vprot_insn,	3,	SUF_Z,	0x01,	0,	0,	0,	CPU_XOP,	0,	0
-vpsadbw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xF6,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsadbw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xF6,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vpshab,	amd_vpshift_insn,	2,	SUF_Z,	0x98,	0,	0,	0,	CPU_XOP,	0,	0
 vpshad,	amd_vpshift_insn,	2,	SUF_Z,	0x9A,	0,	0,	0,	CPU_XOP,	0,	0
 vpshaq,	amd_vpshift_insn,	2,	SUF_Z,	0x9B,	0,	0,	0,	CPU_XOP,	0,	0
@@ -1299,41 +1337,46 @@
 vpshld,	amd_vpshift_insn,	2,	SUF_Z,	0x96,	0,	0,	0,	CPU_XOP,	0,	0
 vpshlq,	amd_vpshift_insn,	2,	SUF_Z,	0x97,	0,	0,	0,	CPU_XOP,	0,	0
 vpshlw,	amd_vpshift_insn,	2,	SUF_Z,	0x95,	0,	0,	0,	CPU_XOP,	0,	0
-vpshufb,	ssse3_insn,	3,	SUF_Z,	0x00,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpshufd,	xmm_xmm128_imm_insn,	1,	SUF_Z,	0x66,	0x70,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpshufhw,	xmm_xmm128_imm_insn,	1,	SUF_Z,	0xF3,	0x70,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpshuflw,	xmm_xmm128_imm_insn,	1,	SUF_Z,	0xF2,	0x70,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsignb,	ssse3_insn,	3,	SUF_Z,	0x08,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsignd,	ssse3_insn,	3,	SUF_Z,	0x0A,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsignw,	ssse3_insn,	3,	SUF_Z,	0x09,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpslld,	vpshift_insn,	4,	SUF_Z,	0xF2,	0x72,	0x06,	ONLY_AVX,	CPU_AVX,	0,	0
-vpslldq,	pslrldq_insn,	2,	SUF_Z,	0x07,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsllq,	vpshift_insn,	4,	SUF_Z,	0xF3,	0x73,	0x06,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsllw,	vpshift_insn,	4,	SUF_Z,	0xF1,	0x71,	0x06,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsrad,	vpshift_insn,	4,	SUF_Z,	0xE2,	0x72,	0x04,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsraw,	vpshift_insn,	4,	SUF_Z,	0xE1,	0x71,	0x04,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsrld,	vpshift_insn,	4,	SUF_Z,	0xD2,	0x72,	0x02,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsrldq,	pslrldq_insn,	2,	SUF_Z,	0x03,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsrlq,	vpshift_insn,	4,	SUF_Z,	0xD3,	0x73,	0x02,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsrlw,	vpshift_insn,	4,	SUF_Z,	0xD1,	0x71,	0x02,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsubb,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xF8,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsubd,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xFA,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsubq,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xFB,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsubsb,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xE8,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsubsw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xE9,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsubusb,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xD8,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsubusw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xD9,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpsubw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xF9,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpshufb,	ssse3_insn,	5,	SUF_Z,	0x00,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpshufd,	xmm_xmm128_imm_256avx2_insn,	2,	SUF_Z,	0x66,	0x70,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpshufhw,	xmm_xmm128_imm_256avx2_insn,	2,	SUF_Z,	0xF3,	0x70,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpshuflw,	xmm_xmm128_imm_256avx2_insn,	2,	SUF_Z,	0xF2,	0x70,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsignb,	ssse3_insn,	5,	SUF_Z,	0x08,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsignd,	ssse3_insn,	5,	SUF_Z,	0x0A,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsignw,	ssse3_insn,	5,	SUF_Z,	0x09,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpslld,	vpshift_insn,	8,	SUF_Z,	0xF2,	0x72,	0x06,	ONLY_AVX,	CPU_AVX,	0,	0
+vpslldq,	pslrldq_insn,	4,	SUF_Z,	0x07,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsllq,	vpshift_insn,	8,	SUF_Z,	0xF3,	0x73,	0x06,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsllvd,	vpshiftv_vexw0_avx2_insn,	2,	SUF_Z,	0x47,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpsllvq,	vpshiftv_vexw1_avx2_insn,	2,	SUF_Z,	0x47,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpsllw,	vpshift_insn,	8,	SUF_Z,	0xF1,	0x71,	0x06,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsrad,	vpshift_insn,	8,	SUF_Z,	0xE2,	0x72,	0x04,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsravd,	vpshiftv_vexw0_avx2_insn,	2,	SUF_Z,	0x46,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpsraw,	vpshift_insn,	8,	SUF_Z,	0xE1,	0x71,	0x04,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsrld,	vpshift_insn,	8,	SUF_Z,	0xD2,	0x72,	0x02,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsrldq,	pslrldq_insn,	4,	SUF_Z,	0x03,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsrlq,	vpshift_insn,	8,	SUF_Z,	0xD3,	0x73,	0x02,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsrlvd,	vpshiftv_vexw0_avx2_insn,	2,	SUF_Z,	0x45,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpsrlvq,	vpshiftv_vexw1_avx2_insn,	2,	SUF_Z,	0x45,	0,	0,	ONLY_AVX,	CPU_AVX2,	0,	0
+vpsrlw,	vpshift_insn,	8,	SUF_Z,	0xD1,	0x71,	0x02,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsubb,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xF8,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsubd,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xFA,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsubq,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xFB,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsubsb,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xE8,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsubsw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xE9,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsubusb,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xD8,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsubusw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xD9,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpsubw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xF9,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vptest,	sse4_insn,	2,	SUF_Z,	0x17,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpunpckhbw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x68,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpunpckhdq,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x6A,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpunpckhqdq,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x6D,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpunpckhwd,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x69,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpunpcklbw,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x60,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpunpckldq,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x62,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpunpcklqdq,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x6C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpunpcklwd,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0x61,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vpxor,	xmm_xmm128_insn,	2,	SUF_Z,	0x66,	0xEF,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpunpckhbw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x68,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpunpckhdq,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x6A,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpunpckhqdq,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x6D,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpunpckhwd,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x69,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpunpcklbw,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x60,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpunpckldq,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x62,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpunpcklqdq,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x6C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpunpcklwd,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0x61,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vpxor,	xmm_xmm128_256avx2_insn,	4,	SUF_Z,	0x66,	0xEF,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vrcpps,	avx_xmm_xmm128_insn,	2,	SUF_Z,	0x00,	0x53,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vrcpss,	xmm_xmm32_insn,	4,	SUF_Z,	0xF3,	0x53,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vroundpd,	avx_sse4imm_insn,	3,	SUF_Z,	0x09,	0,	0,	ONLY_AVX,	CPU_SSE41,	0,	0
@@ -1349,20 +1392,20 @@
 vsqrtsd,	xmm_xmm64_insn,	4,	SUF_Z,	0xF2,	0x51,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vsqrtss,	xmm_xmm32_insn,	4,	SUF_Z,	0xF3,	0x51,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vstmxcsr,	ldstmxcsr_insn,	1,	SUF_Z,	0x03,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vsubpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x5C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vsubps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x00,	0x5C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vsubpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x5C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vsubps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x00,	0x5C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vsubsd,	xmm_xmm64_insn,	4,	SUF_Z,	0xF2,	0x5C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vsubss,	xmm_xmm32_insn,	4,	SUF_Z,	0xF3,	0x5C,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vtestpd,	sse4_insn,	2,	SUF_Z,	0x0F,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vtestps,	sse4_insn,	2,	SUF_Z,	0x0E,	0xC0,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vucomisd,	avx_xmm_xmm64_insn,	2,	SUF_Z,	0x66,	0x2E,	0,	ONLY_AVX,	CPU_AVX,	0,	0
 vucomiss,	avx_xmm_xmm32_insn,	2,	SUF_Z,	0x00,	0x2E,	0,	ONLY_AVX,	CPU_AVX,	0,	0
-vunpckhpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x15,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vunpckhps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x00,	0x15,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vunpcklpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x14,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vunpcklps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x00,	0x14,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vxorpd,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x66,	0x57,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
-vxorps,	xmm_xmm128_256_insn,	3,	SUF_Z,	0x00,	0x57,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vunpckhpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x15,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vunpckhps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x00,	0x15,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vunpcklpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x14,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vunpcklps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x00,	0x14,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vxorpd,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x66,	0x57,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
+vxorps,	xmm_xmm128_256_insn,	4,	SUF_Z,	0x00,	0x57,	0xC0,	ONLY_AVX,	CPU_AVX,	0,	0
 vzeroall,	vzero_insn,	1,	SUF_Z,	0xC4,	0,	0,	0,	CPU_AVX,	0,	0
 vzeroupper,	vzero_insn,	1,	SUF_Z,	0xC0,	0,	0,	0,	CPU_AVX,	0,	0
 wait,	onebyte_insn,	1,	SUF_Z,	0x9B,	0,	0,	0,	0,	0,	0