aboutsummaryrefslogtreecommitdiff
path: root/src/jit/dis_x86.lua
diff options
context:
space:
mode:
Diffstat (limited to 'src/jit/dis_x86.lua')
-rw-r--r--src/jit/dis_x86.lua297
1 files changed, 207 insertions, 90 deletions
diff --git a/src/jit/dis_x86.lua b/src/jit/dis_x86.lua
index 9246820d..364a3184 100644
--- a/src/jit/dis_x86.lua
+++ b/src/jit/dis_x86.lua
@@ -15,19 +15,20 @@
15-- Intel and AMD manuals. The supported instruction set is quite extensive 15-- Intel and AMD manuals. The supported instruction set is quite extensive
16-- and reflects what a current generation Intel or AMD CPU implements in 16-- and reflects what a current generation Intel or AMD CPU implements in
17-- 32 bit and 64 bit mode. Yes, this includes MMX, SSE, SSE2, SSE3, SSSE3, 17-- 32 bit and 64 bit mode. Yes, this includes MMX, SSE, SSE2, SSE3, SSSE3,
18-- SSE4.1, SSE4.2, SSE4a and even privileged and hypervisor (VMX/SVM) 18-- SSE4.1, SSE4.2, SSE4a, AVX, AVX2 and even privileged and hypervisor
19-- instructions. 19-- (VMX/SVM) instructions.
20-- 20--
21-- Notes: 21-- Notes:
22-- * The (useless) a16 prefix, 3DNow and pre-586 opcodes are unsupported. 22-- * The (useless) a16 prefix, 3DNow and pre-586 opcodes are unsupported.
23-- * No attempt at optimization has been made -- it's fast enough for my needs. 23-- * No attempt at optimization has been made -- it's fast enough for my needs.
24-- * The public API may change when more architectures are added.
25------------------------------------------------------------------------------ 24------------------------------------------------------------------------------
26 25
27local type = type 26local type = type
28local sub, byte, format = string.sub, string.byte, string.format 27local sub, byte, format = string.sub, string.byte, string.format
29local match, gmatch, gsub = string.match, string.gmatch, string.gsub 28local match, gmatch, gsub = string.match, string.gmatch, string.gsub
30local lower, rep = string.lower, string.rep 29local lower, rep = string.lower, string.rep
30local bit = require("bit")
31local tohex = bit.tohex
31 32
32-- Map for 1st opcode byte in 32 bit mode. Ugly? Well ... read on. 33-- Map for 1st opcode byte in 32 bit mode. Ugly? Well ... read on.
33local map_opc1_32 = { 34local map_opc1_32 = {
@@ -76,7 +77,7 @@ local map_opc1_32 = {
76"movBRi","movBRi","movBRi","movBRi","movBRi","movBRi","movBRi","movBRi", 77"movBRi","movBRi","movBRi","movBRi","movBRi","movBRi","movBRi","movBRi",
77"movVRI","movVRI","movVRI","movVRI","movVRI","movVRI","movVRI","movVRI", 78"movVRI","movVRI","movVRI","movVRI","movVRI","movVRI","movVRI","movVRI",
78--Cx 79--Cx
79"shift!Bmu","shift!Vmu","retBw","ret","$lesVrm","$ldsVrm","movBmi","movVmi", 80"shift!Bmu","shift!Vmu","retBw","ret","vex*3$lesVrm","vex*2$ldsVrm","movBmi","movVmi",
80"enterBwu","leave","retfBw","retf","int3","intBu","into","iretVS", 81"enterBwu","leave","retfBw","retf","int3","intBu","into","iretVS",
81--Dx 82--Dx
82"shift!Bm1","shift!Vm1","shift!Bmc","shift!Vmc","aamBu","aadBu","salc","xlatb", 83"shift!Bm1","shift!Vm1","shift!Bmc","shift!Vmc","aamBu","aadBu","salc","xlatb",
@@ -101,7 +102,7 @@ local map_opc1_64 = setmetatable({
101 [0x44]="rex*r", [0x45]="rex*rb", [0x46]="rex*rx", [0x47]="rex*rxb", 102 [0x44]="rex*r", [0x45]="rex*rb", [0x46]="rex*rx", [0x47]="rex*rxb",
102 [0x48]="rex*w", [0x49]="rex*wb", [0x4a]="rex*wx", [0x4b]="rex*wxb", 103 [0x48]="rex*w", [0x49]="rex*wb", [0x4a]="rex*wx", [0x4b]="rex*wxb",
103 [0x4c]="rex*wr", [0x4d]="rex*wrb", [0x4e]="rex*wrx", [0x4f]="rex*wrxb", 104 [0x4c]="rex*wr", [0x4d]="rex*wrb", [0x4e]="rex*wrx", [0x4f]="rex*wrxb",
104 [0x82]=false, [0x9a]=false, [0xc4]=false, [0xc5]=false, [0xce]=false, 105 [0x82]=false, [0x9a]=false, [0xc4]="vex*3", [0xc5]="vex*2", [0xce]=false,
105 [0xd4]=false, [0xd5]=false, [0xd6]=false, [0xea]=false, 106 [0xd4]=false, [0xd5]=false, [0xd6]=false, [0xea]=false,
106}, { __index = map_opc1_32 }) 107}, { __index = map_opc1_32 })
107 108
@@ -112,12 +113,12 @@ local map_opc2 = {
112[0]="sldt!Dmp","sgdt!Ump","larVrm","lslVrm",nil,"syscall","clts","sysret", 113[0]="sldt!Dmp","sgdt!Ump","larVrm","lslVrm",nil,"syscall","clts","sysret",
113"invd","wbinvd",nil,"ud1",nil,"$prefetch!Bm","femms","3dnowMrmu", 114"invd","wbinvd",nil,"ud1",nil,"$prefetch!Bm","femms","3dnowMrmu",
114--1x 115--1x
115"movupsXrm|movssXrm|movupdXrm|movsdXrm", 116"movupsXrm|movssXrvm|movupdXrm|movsdXrvm",
116"movupsXmr|movssXmr|movupdXmr|movsdXmr", 117"movupsXmr|movssXmvr|movupdXmr|movsdXmvr",
117"movhlpsXrm$movlpsXrm|movsldupXrm|movlpdXrm|movddupXrm", 118"movhlpsXrm$movlpsXrm|movsldupXrm|movlpdXrm|movddupXrm",
118"movlpsXmr||movlpdXmr", 119"movlpsXmr||movlpdXmr",
119"unpcklpsXrm||unpcklpdXrm", 120"unpcklpsXrvm||unpcklpdXrvm",
120"unpckhpsXrm||unpckhpdXrm", 121"unpckhpsXrvm||unpckhpdXrvm",
121"movlhpsXrm$movhpsXrm|movshdupXrm|movhpdXrm", 122"movlhpsXrm$movhpsXrm|movshdupXrm|movhpdXrm",
122"movhpsXmr||movhpdXmr", 123"movhpsXmr||movhpdXmr",
123"$prefetcht!Bm","hintnopVm","hintnopVm","hintnopVm", 124"$prefetcht!Bm","hintnopVm","hintnopVm","hintnopVm",
@@ -126,7 +127,7 @@ local map_opc2 = {
126"movUmx$","movUmy$","movUxm$","movUym$","movUmz$",nil,"movUzm$",nil, 127"movUmx$","movUmy$","movUxm$","movUym$","movUmz$",nil,"movUzm$",nil,
127"movapsXrm||movapdXrm", 128"movapsXrm||movapdXrm",
128"movapsXmr||movapdXmr", 129"movapsXmr||movapdXmr",
129"cvtpi2psXrMm|cvtsi2ssXrVmt|cvtpi2pdXrMm|cvtsi2sdXrVmt", 130"cvtpi2psXrMm|cvtsi2ssXrvVmt|cvtpi2pdXrMm|cvtsi2sdXrvVmt",
130"movntpsXmr|movntssXmr|movntpdXmr|movntsdXmr", 131"movntpsXmr|movntssXmr|movntpdXmr|movntsdXmr",
131"cvttps2piMrXm|cvttss2siVrXm|cvttpd2piMrXm|cvttsd2siVrXm", 132"cvttps2piMrXm|cvttss2siVrXm|cvttpd2piMrXm|cvttsd2siVrXm",
132"cvtps2piMrXm|cvtss2siVrXm|cvtpd2piMrXm|cvtsd2siVrXm", 133"cvtps2piMrXm|cvtss2siVrXm|cvtpd2piMrXm|cvtsd2siVrXm",
@@ -142,27 +143,27 @@ local map_opc2 = {
142"cmovlVrm","cmovgeVrm","cmovleVrm","cmovgVrm", 143"cmovlVrm","cmovgeVrm","cmovleVrm","cmovgVrm",
143--5x 144--5x
144"movmskpsVrXm$||movmskpdVrXm$","sqrtpsXrm|sqrtssXrm|sqrtpdXrm|sqrtsdXrm", 145"movmskpsVrXm$||movmskpdVrXm$","sqrtpsXrm|sqrtssXrm|sqrtpdXrm|sqrtsdXrm",
145"rsqrtpsXrm|rsqrtssXrm","rcppsXrm|rcpssXrm", 146"rsqrtpsXrm|rsqrtssXrvm","rcppsXrm|rcpssXrvm",
146"andpsXrm||andpdXrm","andnpsXrm||andnpdXrm", 147"andpsXrvm||andpdXrvm","andnpsXrvm||andnpdXrvm",
147"orpsXrm||orpdXrm","xorpsXrm||xorpdXrm", 148"orpsXrvm||orpdXrvm","xorpsXrvm||xorpdXrvm",
148"addpsXrm|addssXrm|addpdXrm|addsdXrm","mulpsXrm|mulssXrm|mulpdXrm|mulsdXrm", 149"addpsXrvm|addssXrvm|addpdXrvm|addsdXrvm","mulpsXrvm|mulssXrvm|mulpdXrvm|mulsdXrvm",
149"cvtps2pdXrm|cvtss2sdXrm|cvtpd2psXrm|cvtsd2ssXrm", 150"cvtps2pdXrm|cvtss2sdXrvm|cvtpd2psXrm|cvtsd2ssXrvm",
150"cvtdq2psXrm|cvttps2dqXrm|cvtps2dqXrm", 151"cvtdq2psXrm|cvttps2dqXrm|cvtps2dqXrm",
151"subpsXrm|subssXrm|subpdXrm|subsdXrm","minpsXrm|minssXrm|minpdXrm|minsdXrm", 152"subpsXrvm|subssXrvm|subpdXrvm|subsdXrvm","minpsXrvm|minssXrvm|minpdXrvm|minsdXrvm",
152"divpsXrm|divssXrm|divpdXrm|divsdXrm","maxpsXrm|maxssXrm|maxpdXrm|maxsdXrm", 153"divpsXrvm|divssXrvm|divpdXrvm|divsdXrvm","maxpsXrvm|maxssXrvm|maxpdXrvm|maxsdXrvm",
153--6x 154--6x
154"punpcklbwPrm","punpcklwdPrm","punpckldqPrm","packsswbPrm", 155"punpcklbwPrvm","punpcklwdPrvm","punpckldqPrvm","packsswbPrvm",
155"pcmpgtbPrm","pcmpgtwPrm","pcmpgtdPrm","packuswbPrm", 156"pcmpgtbPrvm","pcmpgtwPrvm","pcmpgtdPrvm","packuswbPrvm",
156"punpckhbwPrm","punpckhwdPrm","punpckhdqPrm","packssdwPrm", 157"punpckhbwPrvm","punpckhwdPrvm","punpckhdqPrvm","packssdwPrvm",
157"||punpcklqdqXrm","||punpckhqdqXrm", 158"||punpcklqdqXrvm","||punpckhqdqXrvm",
158"movPrVSm","movqMrm|movdquXrm|movdqaXrm", 159"movPrVSm","movqMrm|movdquXrm|movdqaXrm",
159--7x 160--7x
160"pshufwMrmu|pshufhwXrmu|pshufdXrmu|pshuflwXrmu","pshiftw!Pmu", 161"pshufwMrmu|pshufhwXrmu|pshufdXrmu|pshuflwXrmu","pshiftw!Pvmu",
161"pshiftd!Pmu","pshiftq!Mmu||pshiftdq!Xmu", 162"pshiftd!Pvmu","pshiftq!Mvmu||pshiftdq!Xvmu",
162"pcmpeqbPrm","pcmpeqwPrm","pcmpeqdPrm","emms|", 163"pcmpeqbPrvm","pcmpeqwPrvm","pcmpeqdPrvm","emms*|",
163"vmreadUmr||extrqXmuu$|insertqXrmuu$","vmwriteUrm||extrqXrm$|insertqXrm$", 164"vmreadUmr||extrqXmuu$|insertqXrmuu$","vmwriteUrm||extrqXrm$|insertqXrm$",
164nil,nil, 165nil,nil,
165"||haddpdXrm|haddpsXrm","||hsubpdXrm|hsubpsXrm", 166"||haddpdXrvm|haddpsXrvm","||hsubpdXrvm|hsubpsXrvm",
166"movVSmMr|movqXrm|movVSmXr","movqMmr|movdquXmr|movdqaXmr", 167"movVSmMr|movqXrm|movVSmXr","movqMmr|movdquXmr|movdqaXmr",
167--8x 168--8x
168"joVj","jnoVj","jbVj","jnbVj","jzVj","jnzVj","jbeVj","jaVj", 169"joVj","jnoVj","jbVj","jnbVj","jzVj","jnzVj","jbeVj","jaVj",
@@ -180,27 +181,27 @@ nil,nil,
180"bsfVrm","bsrVrm|lzcntVrm|bsrWrm","movsxVrBmt","movsxVrWmt", 181"bsfVrm","bsrVrm|lzcntVrm|bsrWrm","movsxVrBmt","movsxVrWmt",
181--Cx 182--Cx
182"xaddBmr","xaddVmr", 183"xaddBmr","xaddVmr",
183"cmppsXrmu|cmpssXrmu|cmppdXrmu|cmpsdXrmu","$movntiVmr|", 184"cmppsXrvmu|cmpssXrvmu|cmppdXrvmu|cmpsdXrvmu","$movntiVmr|",
184"pinsrwPrWmu","pextrwDrPmu", 185"pinsrwPrvWmu","pextrwDrPmu",
185"shufpsXrmu||shufpdXrmu","$cmpxchg!Qmp", 186"shufpsXrvmu||shufpdXrvmu","$cmpxchg!Qmp",
186"bswapVR","bswapVR","bswapVR","bswapVR","bswapVR","bswapVR","bswapVR","bswapVR", 187"bswapVR","bswapVR","bswapVR","bswapVR","bswapVR","bswapVR","bswapVR","bswapVR",
187--Dx 188--Dx
188"||addsubpdXrm|addsubpsXrm","psrlwPrm","psrldPrm","psrlqPrm", 189"||addsubpdXrvm|addsubpsXrvm","psrlwPrvm","psrldPrvm","psrlqPrvm",
189"paddqPrm","pmullwPrm", 190"paddqPrvm","pmullwPrvm",
190"|movq2dqXrMm|movqXmr|movdq2qMrXm$","pmovmskbVrMm||pmovmskbVrXm", 191"|movq2dqXrMm|movqXmr|movdq2qMrXm$","pmovmskbVrMm||pmovmskbVrXm",
191"psubusbPrm","psubuswPrm","pminubPrm","pandPrm", 192"psubusbPrvm","psubuswPrvm","pminubPrvm","pandPrvm",
192"paddusbPrm","padduswPrm","pmaxubPrm","pandnPrm", 193"paddusbPrvm","padduswPrvm","pmaxubPrvm","pandnPrvm",
193--Ex 194--Ex
194"pavgbPrm","psrawPrm","psradPrm","pavgwPrm", 195"pavgbPrvm","psrawPrvm","psradPrvm","pavgwPrvm",
195"pmulhuwPrm","pmulhwPrm", 196"pmulhuwPrvm","pmulhwPrvm",
196"|cvtdq2pdXrm|cvttpd2dqXrm|cvtpd2dqXrm","$movntqMmr||$movntdqXmr", 197"|cvtdq2pdXrm|cvttpd2dqXrm|cvtpd2dqXrm","$movntqMmr||$movntdqXmr",
197"psubsbPrm","psubswPrm","pminswPrm","porPrm", 198"psubsbPrvm","psubswPrvm","pminswPrvm","porPrvm",
198"paddsbPrm","paddswPrm","pmaxswPrm","pxorPrm", 199"paddsbPrvm","paddswPrvm","pmaxswPrvm","pxorPrvm",
199--Fx 200--Fx
200"|||lddquXrm","psllwPrm","pslldPrm","psllqPrm", 201"|||lddquXrm","psllwPrvm","pslldPrvm","psllqPrvm",
201"pmuludqPrm","pmaddwdPrm","psadbwPrm","maskmovqMrm||maskmovdquXrm$", 202"pmuludqPrvm","pmaddwdPrvm","psadbwPrvm","maskmovqMrm||maskmovdquXrm$",
202"psubbPrm","psubwPrm","psubdPrm","psubqPrm", 203"psubbPrvm","psubwPrvm","psubdPrvm","psubqPrvm",
203"paddbPrm","paddwPrm","padddPrm","ud", 204"paddbPrvm","paddwPrvm","padddPrvm","ud",
204} 205}
205assert(map_opc2[255] == "ud") 206assert(map_opc2[255] == "ud")
206 207
@@ -208,49 +209,91 @@ assert(map_opc2[255] == "ud")
208local map_opc3 = { 209local map_opc3 = {
209["38"] = { -- [66] 0f 38 xx 210["38"] = { -- [66] 0f 38 xx
210--0x 211--0x
211[0]="pshufbPrm","phaddwPrm","phadddPrm","phaddswPrm", 212[0]="pshufbPrvm","phaddwPrvm","phadddPrvm","phaddswPrvm",
212"pmaddubswPrm","phsubwPrm","phsubdPrm","phsubswPrm", 213"pmaddubswPrvm","phsubwPrvm","phsubdPrvm","phsubswPrvm",
213"psignbPrm","psignwPrm","psigndPrm","pmulhrswPrm", 214"psignbPrvm","psignwPrvm","psigndPrvm","pmulhrswPrvm",
214nil,nil,nil,nil, 215"||permilpsXrvm","||permilpdXrvm",nil,nil,
215--1x 216--1x
216"||pblendvbXrma",nil,nil,nil, 217"||pblendvbXrma",nil,nil,nil,
217"||blendvpsXrma","||blendvpdXrma",nil,"||ptestXrm", 218"||blendvpsXrma","||blendvpdXrma","||permpsXrvm","||ptestXrm",
218nil,nil,nil,nil, 219"||broadcastssXrm","||broadcastsdXrm","||broadcastf128XrlXm",nil,
219"pabsbPrm","pabswPrm","pabsdPrm",nil, 220"pabsbPrm","pabswPrm","pabsdPrm",nil,
220--2x 221--2x
221"||pmovsxbwXrm","||pmovsxbdXrm","||pmovsxbqXrm","||pmovsxwdXrm", 222"||pmovsxbwXrm","||pmovsxbdXrm","||pmovsxbqXrm","||pmovsxwdXrm",
222"||pmovsxwqXrm","||pmovsxdqXrm",nil,nil, 223"||pmovsxwqXrm","||pmovsxdqXrm",nil,nil,
223"||pmuldqXrm","||pcmpeqqXrm","||$movntdqaXrm","||packusdwXrm", 224"||pmuldqXrvm","||pcmpeqqXrvm","||$movntdqaXrm","||packusdwXrvm",
224nil,nil,nil,nil, 225"||maskmovpsXrvm","||maskmovpdXrvm","||maskmovpsXmvr","||maskmovpdXmvr",
225--3x 226--3x
226"||pmovzxbwXrm","||pmovzxbdXrm","||pmovzxbqXrm","||pmovzxwdXrm", 227"||pmovzxbwXrm","||pmovzxbdXrm","||pmovzxbqXrm","||pmovzxwdXrm",
227"||pmovzxwqXrm","||pmovzxdqXrm",nil,"||pcmpgtqXrm", 228"||pmovzxwqXrm","||pmovzxdqXrm","||permdXrvm","||pcmpgtqXrvm",
228"||pminsbXrm","||pminsdXrm","||pminuwXrm","||pminudXrm", 229"||pminsbXrvm","||pminsdXrvm","||pminuwXrvm","||pminudXrvm",
229"||pmaxsbXrm","||pmaxsdXrm","||pmaxuwXrm","||pmaxudXrm", 230"||pmaxsbXrvm","||pmaxsdXrvm","||pmaxuwXrvm","||pmaxudXrvm",
230--4x 231--4x
231"||pmulddXrm","||phminposuwXrm", 232"||pmulddXrvm","||phminposuwXrm",nil,nil,
233nil,"||psrlvVSXrvm","||psravdXrvm","||psllvVSXrvm",
234--5x
235[0x58] = "||pbroadcastdXrlXm",[0x59] = "||pbroadcastqXrlXm",
236[0x5a] = "||broadcasti128XrlXm",
237--7x
238[0x78] = "||pbroadcastbXrlXm",[0x79] = "||pbroadcastwXrlXm",
239--8x
240[0x8c] = "||pmaskmovXrvVSm",
241[0x8e] = "||pmaskmovVSmXvr",
242--9x
243[0x96] = "||fmaddsub132pHXrvm",[0x97] = "||fmsubadd132pHXrvm",
244[0x98] = "||fmadd132pHXrvm",[0x99] = "||fmadd132sHXrvm",
245[0x9a] = "||fmsub132pHXrvm",[0x9b] = "||fmsub132sHXrvm",
246[0x9c] = "||fnmadd132pHXrvm",[0x9d] = "||fnmadd132sHXrvm",
247[0x9e] = "||fnmsub132pHXrvm",[0x9f] = "||fnmsub132sHXrvm",
248--Ax
249[0xa6] = "||fmaddsub213pHXrvm",[0xa7] = "||fmsubadd213pHXrvm",
250[0xa8] = "||fmadd213pHXrvm",[0xa9] = "||fmadd213sHXrvm",
251[0xaa] = "||fmsub213pHXrvm",[0xab] = "||fmsub213sHXrvm",
252[0xac] = "||fnmadd213pHXrvm",[0xad] = "||fnmadd213sHXrvm",
253[0xae] = "||fnmsub213pHXrvm",[0xaf] = "||fnmsub213sHXrvm",
254--Bx
255[0xb6] = "||fmaddsub231pHXrvm",[0xb7] = "||fmsubadd231pHXrvm",
256[0xb8] = "||fmadd231pHXrvm",[0xb9] = "||fmadd231sHXrvm",
257[0xba] = "||fmsub231pHXrvm",[0xbb] = "||fmsub231sHXrvm",
258[0xbc] = "||fnmadd231pHXrvm",[0xbd] = "||fnmadd231sHXrvm",
259[0xbe] = "||fnmsub231pHXrvm",[0xbf] = "||fnmsub231sHXrvm",
260--Dx
261[0xdc] = "||aesencXrvm", [0xdd] = "||aesenclastXrvm",
262[0xde] = "||aesdecXrvm", [0xdf] = "||aesdeclastXrvm",
232--Fx 263--Fx
233[0xf0] = "|||crc32TrBmt",[0xf1] = "|||crc32TrVmt", 264[0xf0] = "|||crc32TrBmt",[0xf1] = "|||crc32TrVmt",
265[0xf7] = "| sarxVrmv| shlxVrmv| shrxVrmv",
234}, 266},
235 267
236["3a"] = { -- [66] 0f 3a xx 268["3a"] = { -- [66] 0f 3a xx
237--0x 269--0x
238[0x00]=nil,nil,nil,nil,nil,nil,nil,nil, 270[0x00]="||permqXrmu","||permpdXrmu","||pblenddXrvmu",nil,
239"||roundpsXrmu","||roundpdXrmu","||roundssXrmu","||roundsdXrmu", 271"||permilpsXrmu","||permilpdXrmu","||perm2f128Xrvmu",nil,
240"||blendpsXrmu","||blendpdXrmu","||pblendwXrmu","palignrPrmu", 272"||roundpsXrmu","||roundpdXrmu","||roundssXrvmu","||roundsdXrvmu",
273"||blendpsXrvmu","||blendpdXrvmu","||pblendwXrvmu","palignrPrvmu",
241--1x 274--1x
242nil,nil,nil,nil, 275nil,nil,nil,nil,
243"||pextrbVmXru","||pextrwVmXru","||pextrVmSXru","||extractpsVmXru", 276"||pextrbVmXru","||pextrwVmXru","||pextrVmSXru","||extractpsVmXru",
244nil,nil,nil,nil,nil,nil,nil,nil, 277"||insertf128XrvlXmu","||extractf128XlXmYru",nil,nil,
278nil,nil,nil,nil,
245--2x 279--2x
246"||pinsrbXrVmu","||insertpsXrmu","||pinsrXrVmuS",nil, 280"||pinsrbXrvVmu","||insertpsXrvmu","||pinsrXrvVmuS",nil,
281--3x
282[0x38] = "||inserti128Xrvmu",[0x39] = "||extracti128XlXmYru",
247--4x 283--4x
248[0x40] = "||dppsXrmu", 284[0x40] = "||dppsXrvmu",
249[0x41] = "||dppdXrmu", 285[0x41] = "||dppdXrvmu",
250[0x42] = "||mpsadbwXrmu", 286[0x42] = "||mpsadbwXrvmu",
287[0x44] = "||pclmulqdqXrvmu",
288[0x46] = "||perm2i128Xrvmu",
289[0x4a] = "||blendvpsXrvmb",[0x4b] = "||blendvpdXrvmb",
290[0x4c] = "||pblendvbXrvmb",
251--6x 291--6x
252[0x60] = "||pcmpestrmXrmu",[0x61] = "||pcmpestriXrmu", 292[0x60] = "||pcmpestrmXrmu",[0x61] = "||pcmpestriXrmu",
253[0x62] = "||pcmpistrmXrmu",[0x63] = "||pcmpistriXrmu", 293[0x62] = "||pcmpistrmXrmu",[0x63] = "||pcmpistriXrmu",
294[0xdf] = "||aeskeygenassistXrmu",
295--Fx
296[0xf0] = "||| rorxVrmu",
254}, 297},
255} 298}
256 299
@@ -354,17 +397,19 @@ local map_regs = {
354 "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7" }, -- No x64 ext! 397 "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7" }, -- No x64 ext!
355 X = { "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7", 398 X = { "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
356 "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15" }, 399 "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15" },
400 Y = { "ymm0", "ymm1", "ymm2", "ymm3", "ymm4", "ymm5", "ymm6", "ymm7",
401 "ymm8", "ymm9", "ymm10", "ymm11", "ymm12", "ymm13", "ymm14", "ymm15" },
357} 402}
358local map_segregs = { "es", "cs", "ss", "ds", "fs", "gs", "segr6", "segr7" } 403local map_segregs = { "es", "cs", "ss", "ds", "fs", "gs", "segr6", "segr7" }
359 404
360-- Maps for size names. 405-- Maps for size names.
361local map_sz2n = { 406local map_sz2n = {
362 B = 1, W = 2, D = 4, Q = 8, M = 8, X = 16, 407 B = 1, W = 2, D = 4, Q = 8, M = 8, X = 16, Y = 32,
363} 408}
364local map_sz2prefix = { 409local map_sz2prefix = {
365 B = "byte", W = "word", D = "dword", 410 B = "byte", W = "word", D = "dword",
366 Q = "qword", 411 Q = "qword",
367 M = "qword", X = "xword", 412 M = "qword", X = "xword", Y = "yword",
368 F = "dword", G = "qword", -- No need for sizes/register names for these two. 413 F = "dword", G = "qword", -- No need for sizes/register names for these two.
369} 414}
370 415
@@ -387,10 +432,13 @@ local function putop(ctx, text, operands)
387 if ctx.rep then text = ctx.rep.." "..text; ctx.rep = false end 432 if ctx.rep then text = ctx.rep.." "..text; ctx.rep = false end
388 if ctx.rex then 433 if ctx.rex then
389 local t = (ctx.rexw and "w" or "")..(ctx.rexr and "r" or "").. 434 local t = (ctx.rexw and "w" or "")..(ctx.rexr and "r" or "")..
390 (ctx.rexx and "x" or "")..(ctx.rexb and "b" or "") 435 (ctx.rexx and "x" or "")..(ctx.rexb and "b" or "")..
391 if t ~= "" then text = "rex."..t.." "..text end 436 (ctx.vexl and "l" or "")
437 if ctx.vexv and ctx.vexv ~= 0 then t = t.."v"..ctx.vexv end
438 if t ~= "" then text = ctx.rex.."."..t.." "..gsub(text, "^ ", "")
439 elseif ctx.rex == "vex" then text = gsub("v"..text, "^v ", "") end
392 ctx.rexw = false; ctx.rexr = false; ctx.rexx = false; ctx.rexb = false 440 ctx.rexw = false; ctx.rexr = false; ctx.rexx = false; ctx.rexb = false
393 ctx.rex = false 441 ctx.rex = false; ctx.vexl = false; ctx.vexv = false
394 end 442 end
395 if ctx.seg then 443 if ctx.seg then
396 local text2, n = gsub(text, "%[", "["..ctx.seg..":") 444 local text2, n = gsub(text, "%[", "["..ctx.seg..":")
@@ -405,6 +453,7 @@ local function putop(ctx, text, operands)
405 end 453 end
406 ctx.out(format("%08x %s%s\n", ctx.addr+ctx.start, hex, text)) 454 ctx.out(format("%08x %s%s\n", ctx.addr+ctx.start, hex, text))
407 ctx.mrm = false 455 ctx.mrm = false
456 ctx.vexv = false
408 ctx.start = pos 457 ctx.start = pos
409 ctx.imm = nil 458 ctx.imm = nil
410end 459end
@@ -413,7 +462,7 @@ end
413local function clearprefixes(ctx) 462local function clearprefixes(ctx)
414 ctx.o16 = false; ctx.seg = false; ctx.lock = false; ctx.rep = false 463 ctx.o16 = false; ctx.seg = false; ctx.lock = false; ctx.rep = false
415 ctx.rexw = false; ctx.rexr = false; ctx.rexx = false; ctx.rexb = false 464 ctx.rexw = false; ctx.rexr = false; ctx.rexx = false; ctx.rexb = false
416 ctx.rex = false; ctx.a32 = false 465 ctx.rex = false; ctx.a32 = false; ctx.vexl = false
417end 466end
418 467
419-- Fallback for incomplete opcodes at the end. 468-- Fallback for incomplete opcodes at the end.
@@ -450,9 +499,9 @@ end
450-- Process pattern string and generate the operands. 499-- Process pattern string and generate the operands.
451local function putpat(ctx, name, pat) 500local function putpat(ctx, name, pat)
452 local operands, regs, sz, mode, sp, rm, sc, rx, sdisp 501 local operands, regs, sz, mode, sp, rm, sc, rx, sdisp
453 local code, pos, stop = ctx.code, ctx.pos, ctx.stop 502 local code, pos, stop, vexl = ctx.code, ctx.pos, ctx.stop, ctx.vexl
454 503
455 -- Chars used: 1DFGIMPQRSTUVWXacdfgijmoprstuwxyz 504 -- Chars used: 1DFGHIMPQRSTUVWXYabcdfgijlmoprstuvwxyz
456 for p in gmatch(pat, ".") do 505 for p in gmatch(pat, ".") do
457 local x = nil 506 local x = nil
458 if p == "V" or p == "U" then 507 if p == "V" or p == "U" then
@@ -467,12 +516,17 @@ local function putpat(ctx, name, pat)
467 elseif p == "B" then 516 elseif p == "B" then
468 sz = "B" 517 sz = "B"
469 regs = ctx.rex and map_regs.B64 or map_regs.B 518 regs = ctx.rex and map_regs.B64 or map_regs.B
470 elseif match(p, "[WDQMXFG]") then 519 elseif match(p, "[WDQMXYFG]") then
471 sz = p 520 sz = p
521 if sz == "X" and vexl then sz = "Y"; ctx.vexl = false end
472 regs = map_regs[sz] 522 regs = map_regs[sz]
473 elseif p == "P" then 523 elseif p == "P" then
474 sz = ctx.o16 and "X" or "M"; ctx.o16 = false 524 sz = ctx.o16 and "X" or "M"; ctx.o16 = false
525 if sz == "X" and vexl then sz = "Y"; ctx.vexl = false end
475 regs = map_regs[sz] 526 regs = map_regs[sz]
527 elseif p == "H" then
528 name = name..(ctx.rexw and "d" or "s")
529 ctx.rexw = false
476 elseif p == "S" then 530 elseif p == "S" then
477 name = name..lower(sz) 531 name = name..lower(sz)
478 elseif p == "s" then 532 elseif p == "s" then
@@ -484,6 +538,10 @@ local function putpat(ctx, name, pat)
484 local imm = getimm(ctx, pos, 1); if not imm then return end 538 local imm = getimm(ctx, pos, 1); if not imm then return end
485 x = format("0x%02x", imm) 539 x = format("0x%02x", imm)
486 pos = pos+1 540 pos = pos+1
541 elseif p == "b" then
542 local imm = getimm(ctx, pos, 1); if not imm then return end
543 x = regs[imm/16+1]
544 pos = pos+1
487 elseif p == "w" then 545 elseif p == "w" then
488 local imm = getimm(ctx, pos, 2); if not imm then return end 546 local imm = getimm(ctx, pos, 2); if not imm then return end
489 x = format("0x%x", imm) 547 x = format("0x%x", imm)
@@ -532,7 +590,7 @@ local function putpat(ctx, name, pat)
532 local lo = imm % 0x1000000 590 local lo = imm % 0x1000000
533 x = format("0x%02x%06x", (imm-lo) / 0x1000000, lo) 591 x = format("0x%02x%06x", (imm-lo) / 0x1000000, lo)
534 else 592 else
535 x = format("0x%08x", imm) 593 x = "0x"..tohex(imm)
536 end 594 end
537 elseif p == "R" then 595 elseif p == "R" then
538 local r = byte(code, pos-1, pos-1)%8 596 local r = byte(code, pos-1, pos-1)%8
@@ -616,8 +674,13 @@ local function putpat(ctx, name, pat)
616 else 674 else
617 x = "CR"..sp 675 x = "CR"..sp
618 end 676 end
677 elseif p == "v" then
678 if ctx.vexv then
679 x = regs[ctx.vexv+1]; ctx.vexv = false
680 end
619 elseif p == "y" then x = "DR"..sp 681 elseif p == "y" then x = "DR"..sp
620 elseif p == "z" then x = "TR"..sp 682 elseif p == "z" then x = "TR"..sp
683 elseif p == "l" then vexl = false
621 elseif p == "t" then 684 elseif p == "t" then
622 else 685 else
623 error("bad pattern `"..pat.."'") 686 error("bad pattern `"..pat.."'")
@@ -692,7 +755,8 @@ map_act = {
692 B = putpat, W = putpat, D = putpat, Q = putpat, 755 B = putpat, W = putpat, D = putpat, Q = putpat,
693 V = putpat, U = putpat, T = putpat, 756 V = putpat, U = putpat, T = putpat,
694 M = putpat, X = putpat, P = putpat, 757 M = putpat, X = putpat, P = putpat,
695 F = putpat, G = putpat, 758 F = putpat, G = putpat, Y = putpat,
759 H = putpat,
696 760
697 -- Collect prefixes. 761 -- Collect prefixes.
698 [":"] = function(ctx, name, pat) 762 [":"] = function(ctx, name, pat)
@@ -753,15 +817,68 @@ map_act = {
753 817
754 -- REX prefix. 818 -- REX prefix.
755 rex = function(ctx, name, pat) 819 rex = function(ctx, name, pat)
756 if ctx.rex then return unknown(ctx) end -- Only 1 REX prefix allowed. 820 if ctx.rex then return unknown(ctx) end -- Only 1 REX or VEX prefix allowed.
757 for p in gmatch(pat, ".") do ctx["rex"..p] = true end 821 for p in gmatch(pat, ".") do ctx["rex"..p] = true end
758 ctx.rex = true 822 ctx.rex = "rex"
823 end,
824
825 -- VEX prefix.
826 vex = function(ctx, name, pat)
827 if ctx.rex then return unknown(ctx) end -- Only 1 REX or VEX prefix allowed.
828 ctx.rex = "vex"
829 local pos = ctx.pos
830 if ctx.mrm then
831 ctx.mrm = nil
832 pos = pos-1
833 end
834 local b = byte(ctx.code, pos, pos)
835 if not b then return incomplete(ctx) end
836 pos = pos+1
837 if b < 128 then ctx.rexr = true end
838 local m = 1
839 if pat == "3" then
840 m = b%32; b = (b-m)/32
841 local nb = b%2; b = (b-nb)/2
842 if nb == 0 then ctx.rexb = true end
843 local nx = b%2
844 if nx == 0 then ctx.rexx = true end
845 b = byte(ctx.code, pos, pos)
846 if not b then return incomplete(ctx) end
847 pos = pos+1
848 if b >= 128 then ctx.rexw = true end
849 end
850 ctx.pos = pos
851 local map
852 if m == 1 then map = map_opc2
853 elseif m == 2 then map = map_opc3["38"]
854 elseif m == 3 then map = map_opc3["3a"]
855 else return unknown(ctx) end
856 local p = b%4; b = (b-p)/4
857 if p == 1 then ctx.o16 = "o16"
858 elseif p == 2 then ctx.rep = "rep"
859 elseif p == 3 then ctx.rep = "repne" end
860 local l = b%2; b = (b-l)/2
861 if l ~= 0 then ctx.vexl = true end
862 ctx.vexv = (-1-b)%16
863 return dispatchmap(ctx, map)
759 end, 864 end,
760 865
761 -- Special case for nop with REX prefix. 866 -- Special case for nop with REX prefix.
762 nop = function(ctx, name, pat) 867 nop = function(ctx, name, pat)
763 return dispatch(ctx, ctx.rex and pat or "nop") 868 return dispatch(ctx, ctx.rex and pat or "nop")
764 end, 869 end,
870
871 -- Special case for 0F 77.
872 emms = function(ctx, name, pat)
873 if ctx.rex ~= "vex" then
874 return putop(ctx, "emms")
875 elseif ctx.vexl then
876 ctx.vexl = false
877 return putop(ctx, "zeroall")
878 else
879 return putop(ctx, "zeroupper")
880 end
881 end,
765} 882}
766 883
767------------------------------------------------------------------------------ 884------------------------------------------------------------------------------
@@ -782,7 +899,7 @@ local function disass_block(ctx, ofs, len)
782end 899end
783 900
784-- Extended API: create a disassembler context. Then call ctx:disass(ofs, len). 901-- Extended API: create a disassembler context. Then call ctx:disass(ofs, len).
785local function create_(code, addr, out) 902local function create(code, addr, out)
786 local ctx = {} 903 local ctx = {}
787 ctx.code = code 904 ctx.code = code
788 ctx.addr = (addr or 0) - 1 905 ctx.addr = (addr or 0) - 1
@@ -796,8 +913,8 @@ local function create_(code, addr, out)
796 return ctx 913 return ctx
797end 914end
798 915
799local function create64_(code, addr, out) 916local function create64(code, addr, out)
800 local ctx = create_(code, addr, out) 917 local ctx = create(code, addr, out)
801 ctx.x64 = true 918 ctx.x64 = true
802 ctx.map1 = map_opc1_64 919 ctx.map1 = map_opc1_64
803 ctx.aregs = map_regs.Q 920 ctx.aregs = map_regs.Q
@@ -805,32 +922,32 @@ local function create64_(code, addr, out)
805end 922end
806 923
807-- Simple API: disassemble code (a string) at address and output via out. 924-- Simple API: disassemble code (a string) at address and output via out.
808local function disass_(code, addr, out) 925local function disass(code, addr, out)
809 create_(code, addr, out):disass() 926 create(code, addr, out):disass()
810end 927end
811 928
812local function disass64_(code, addr, out) 929local function disass64(code, addr, out)
813 create64_(code, addr, out):disass() 930 create64(code, addr, out):disass()
814end 931end
815 932
816-- Return register name for RID. 933-- Return register name for RID.
817local function regname_(r) 934local function regname(r)
818 if r < 8 then return map_regs.D[r+1] end 935 if r < 8 then return map_regs.D[r+1] end
819 return map_regs.X[r-7] 936 return map_regs.X[r-7]
820end 937end
821 938
822local function regname64_(r) 939local function regname64(r)
823 if r < 16 then return map_regs.Q[r+1] end 940 if r < 16 then return map_regs.Q[r+1] end
824 return map_regs.X[r-15] 941 return map_regs.X[r-15]
825end 942end
826 943
827-- Public module functions. 944-- Public module functions.
828module(...) 945return {
829 946 create = create,
830create = create_ 947 create64 = create64,
831create64 = create64_ 948 disass = disass,
832disass = disass_ 949 disass64 = disass64,
833disass64 = disass64_ 950 regname = regname,
834regname = regname_ 951 regname64 = regname64
835regname64 = regname64_ 952}
836 953