diff --git a/polymer/eduke32/Makefile b/polymer/eduke32/Makefile index 2aee2bb23..baee08a3e 100644 --- a/polymer/eduke32/Makefile +++ b/polymer/eduke32/Makefile @@ -159,6 +159,10 @@ ifneq (0,$(LUNATIC)) $(OBJ)/luaJIT_BC_xmath.$o \ $(OBJ)/luaJIT_BC_defs.$o \ $(OBJ)/luaJIT_BC_v.$o \ + $(OBJ)/luaJIT_BC_dump.$o \ + $(OBJ)/luaJIT_BC_dis_x86.$o \ + $(OBJ)/luaJIT_BC_dis_x64.$o \ + # TODO: remove debugging modules from release build # now, take care of having the necessary symbols (sector, wall, etc.) in the # executable no matter what the debugging level diff --git a/polymer/eduke32/source/lunatic/defs_common.lua b/polymer/eduke32/source/lunatic/defs_common.lua index 13ff4c43e..bcb6844eb 100644 --- a/polymer/eduke32/source/lunatic/defs_common.lua +++ b/polymer/eduke32/source/lunatic/defs_common.lua @@ -12,13 +12,16 @@ local bit = require("bit") -- ~=0: print diagnostic information -- 2: disable JIT compilation -- 4: load LuaJIT's 'v' module, printing trace info +-- 8: load LuaJIT's 'dump' module, printing generated IR/machine code ffi.cdef "enum { _DEBUG_LUNATIC=1 }" if (bit.band(ffiC._DEBUG_LUNATIC, 2)~=0) then require("jit").off() end -if (bit.band(ffiC._DEBUG_LUNATIC, 4)~=0) then +if (bit.band(ffiC._DEBUG_LUNATIC, 8)~=0) then + require("dump").on("T") -- raw text output +elseif (bit.band(ffiC._DEBUG_LUNATIC, 4)~=0) then require("v").on() end diff --git a/polymer/eduke32/source/lunatic/dis_x64.lua b/polymer/eduke32/source/lunatic/dis_x64.lua new file mode 100644 index 000000000..4a1894ace --- /dev/null +++ b/polymer/eduke32/source/lunatic/dis_x64.lua @@ -0,0 +1,20 @@ +---------------------------------------------------------------------------- +-- LuaJIT x64 disassembler wrapper module. +-- +-- Copyright (C) 2005-2013 Mike Pall. All rights reserved. +-- Released under the MIT license. See Copyright Notice in luajit.h +---------------------------------------------------------------------------- +-- This module just exports the 64 bit functions from the combined +-- x86/x64 disassembler module. All the interesting stuff is there. +------------------------------------------------------------------------------ + +local require = require + +module(...) + +local dis_x86 = require(_PACKAGE.."dis_x86") + +create = dis_x86.create64 +disass = dis_x86.disass64 +regname = dis_x86.regname64 + diff --git a/polymer/eduke32/source/lunatic/dis_x86.lua b/polymer/eduke32/source/lunatic/dis_x86.lua new file mode 100644 index 000000000..c442a176b --- /dev/null +++ b/polymer/eduke32/source/lunatic/dis_x86.lua @@ -0,0 +1,836 @@ +---------------------------------------------------------------------------- +-- LuaJIT x86/x64 disassembler module. +-- +-- Copyright (C) 2005-2013 Mike Pall. All rights reserved. +-- Released under the MIT license. See Copyright Notice in luajit.h +---------------------------------------------------------------------------- +-- This is a helper module used by the LuaJIT machine code dumper module. +-- +-- Sending small code snippets to an external disassembler and mixing the +-- output with our own stuff was too fragile. So I had to bite the bullet +-- and write yet another x86 disassembler. Oh well ... +-- +-- The output format is very similar to what ndisasm generates. But it has +-- been developed independently by looking at the opcode tables from the +-- Intel and AMD manuals. The supported instruction set is quite extensive +-- and reflects what a current generation Intel or AMD CPU implements in +-- 32 bit and 64 bit mode. Yes, this includes MMX, SSE, SSE2, SSE3, SSSE3, +-- SSE4.1, SSE4.2, SSE4a and even privileged and hypervisor (VMX/SVM) +-- instructions. +-- +-- Notes: +-- * The (useless) a16 prefix, 3DNow and pre-586 opcodes are unsupported. +-- * No attempt at optimization has been made -- it's fast enough for my needs. +-- * The public API may change when more architectures are added. +------------------------------------------------------------------------------ + +local type = type +local sub, byte, format = string.sub, string.byte, string.format +local match, gmatch, gsub = string.match, string.gmatch, string.gsub +local lower, rep = string.lower, string.rep + +-- Map for 1st opcode byte in 32 bit mode. Ugly? Well ... read on. +local map_opc1_32 = { +--0x +[0]="addBmr","addVmr","addBrm","addVrm","addBai","addVai","push es","pop es", +"orBmr","orVmr","orBrm","orVrm","orBai","orVai","push cs","opc2*", +--1x +"adcBmr","adcVmr","adcBrm","adcVrm","adcBai","adcVai","push ss","pop ss", +"sbbBmr","sbbVmr","sbbBrm","sbbVrm","sbbBai","sbbVai","push ds","pop ds", +--2x +"andBmr","andVmr","andBrm","andVrm","andBai","andVai","es:seg","daa", +"subBmr","subVmr","subBrm","subVrm","subBai","subVai","cs:seg","das", +--3x +"xorBmr","xorVmr","xorBrm","xorVrm","xorBai","xorVai","ss:seg","aaa", +"cmpBmr","cmpVmr","cmpBrm","cmpVrm","cmpBai","cmpVai","ds:seg","aas", +--4x +"incVR","incVR","incVR","incVR","incVR","incVR","incVR","incVR", +"decVR","decVR","decVR","decVR","decVR","decVR","decVR","decVR", +--5x +"pushUR","pushUR","pushUR","pushUR","pushUR","pushUR","pushUR","pushUR", +"popUR","popUR","popUR","popUR","popUR","popUR","popUR","popUR", +--6x +"sz*pushaw,pusha","sz*popaw,popa","boundVrm","arplWmr", +"fs:seg","gs:seg","o16:","a16", +"pushUi","imulVrmi","pushBs","imulVrms", +"insb","insVS","outsb","outsVS", +--7x +"joBj","jnoBj","jbBj","jnbBj","jzBj","jnzBj","jbeBj","jaBj", +"jsBj","jnsBj","jpeBj","jpoBj","jlBj","jgeBj","jleBj","jgBj", +--8x +"arith!Bmi","arith!Vmi","arith!Bmi","arith!Vms", +"testBmr","testVmr","xchgBrm","xchgVrm", +"movBmr","movVmr","movBrm","movVrm", +"movVmg","leaVrm","movWgm","popUm", +--9x +"nop*xchgVaR|pause|xchgWaR|repne nop","xchgVaR","xchgVaR","xchgVaR", +"xchgVaR","xchgVaR","xchgVaR","xchgVaR", +"sz*cbw,cwde,cdqe","sz*cwd,cdq,cqo","call farViw","wait", +"sz*pushfw,pushf","sz*popfw,popf","sahf","lahf", +--Ax +"movBao","movVao","movBoa","movVoa", +"movsb","movsVS","cmpsb","cmpsVS", +"testBai","testVai","stosb","stosVS", +"lodsb","lodsVS","scasb","scasVS", +--Bx +"movBRi","movBRi","movBRi","movBRi","movBRi","movBRi","movBRi","movBRi", +"movVRI","movVRI","movVRI","movVRI","movVRI","movVRI","movVRI","movVRI", +--Cx +"shift!Bmu","shift!Vmu","retBw","ret","$lesVrm","$ldsVrm","movBmi","movVmi", +"enterBwu","leave","retfBw","retf","int3","intBu","into","iretVS", +--Dx +"shift!Bm1","shift!Vm1","shift!Bmc","shift!Vmc","aamBu","aadBu","salc","xlatb", +"fp*0","fp*1","fp*2","fp*3","fp*4","fp*5","fp*6","fp*7", +--Ex +"loopneBj","loopeBj","loopBj","sz*jcxzBj,jecxzBj,jrcxzBj", +"inBau","inVau","outBua","outVua", +"callVj","jmpVj","jmp farViw","jmpBj","inBad","inVad","outBda","outVda", +--Fx +"lock:","int1","repne:rep","rep:","hlt","cmc","testb!Bm","testv!Vm", +"clc","stc","cli","sti","cld","std","incb!Bm","incd!Vm", +} +assert(#map_opc1_32 == 255) + +-- Map for 1st opcode byte in 64 bit mode (overrides only). +local map_opc1_64 = setmetatable({ + [0x06]=false, [0x07]=false, [0x0e]=false, + [0x16]=false, [0x17]=false, [0x1e]=false, [0x1f]=false, + [0x27]=false, [0x2f]=false, [0x37]=false, [0x3f]=false, + [0x60]=false, [0x61]=false, [0x62]=false, [0x63]="movsxdVrDmt", [0x67]="a32:", + [0x40]="rex*", [0x41]="rex*b", [0x42]="rex*x", [0x43]="rex*xb", + [0x44]="rex*r", [0x45]="rex*rb", [0x46]="rex*rx", [0x47]="rex*rxb", + [0x48]="rex*w", [0x49]="rex*wb", [0x4a]="rex*wx", [0x4b]="rex*wxb", + [0x4c]="rex*wr", [0x4d]="rex*wrb", [0x4e]="rex*wrx", [0x4f]="rex*wrxb", + [0x82]=false, [0x9a]=false, [0xc4]=false, [0xc5]=false, [0xce]=false, + [0xd4]=false, [0xd5]=false, [0xd6]=false, [0xea]=false, +}, { __index = map_opc1_32 }) + +-- Map for 2nd opcode byte (0F xx). True CISC hell. Hey, I told you. +-- Prefix dependent MMX/SSE opcodes: (none)|rep|o16|repne, -|F3|66|F2 +local map_opc2 = { +--0x +[0]="sldt!Dmp","sgdt!Ump","larVrm","lslVrm",nil,"syscall","clts","sysret", +"invd","wbinvd",nil,"ud1",nil,"$prefetch!Bm","femms","3dnowMrmu", +--1x +"movupsXrm|movssXrm|movupdXrm|movsdXrm", +"movupsXmr|movssXmr|movupdXmr|movsdXmr", +"movhlpsXrm$movlpsXrm|movsldupXrm|movlpdXrm|movddupXrm", +"movlpsXmr||movlpdXmr", +"unpcklpsXrm||unpcklpdXrm", +"unpckhpsXrm||unpckhpdXrm", +"movlhpsXrm$movhpsXrm|movshdupXrm|movhpdXrm", +"movhpsXmr||movhpdXmr", +"$prefetcht!Bm","hintnopVm","hintnopVm","hintnopVm", +"hintnopVm","hintnopVm","hintnopVm","hintnopVm", +--2x +"movUmx$","movUmy$","movUxm$","movUym$","movUmz$",nil,"movUzm$",nil, +"movapsXrm||movapdXrm", +"movapsXmr||movapdXmr", +"cvtpi2psXrMm|cvtsi2ssXrVmt|cvtpi2pdXrMm|cvtsi2sdXrVmt", +"movntpsXmr|movntssXmr|movntpdXmr|movntsdXmr", +"cvttps2piMrXm|cvttss2siVrXm|cvttpd2piMrXm|cvttsd2siVrXm", +"cvtps2piMrXm|cvtss2siVrXm|cvtpd2piMrXm|cvtsd2siVrXm", +"ucomissXrm||ucomisdXrm", +"comissXrm||comisdXrm", +--3x +"wrmsr","rdtsc","rdmsr","rdpmc","sysenter","sysexit",nil,"getsec", +"opc3*38",nil,"opc3*3a",nil,nil,nil,nil,nil, +--4x +"cmovoVrm","cmovnoVrm","cmovbVrm","cmovnbVrm", +"cmovzVrm","cmovnzVrm","cmovbeVrm","cmovaVrm", +"cmovsVrm","cmovnsVrm","cmovpeVrm","cmovpoVrm", +"cmovlVrm","cmovgeVrm","cmovleVrm","cmovgVrm", +--5x +"movmskpsVrXm$||movmskpdVrXm$","sqrtpsXrm|sqrtssXrm|sqrtpdXrm|sqrtsdXrm", +"rsqrtpsXrm|rsqrtssXrm","rcppsXrm|rcpssXrm", +"andpsXrm||andpdXrm","andnpsXrm||andnpdXrm", +"orpsXrm||orpdXrm","xorpsXrm||xorpdXrm", +"addpsXrm|addssXrm|addpdXrm|addsdXrm","mulpsXrm|mulssXrm|mulpdXrm|mulsdXrm", +"cvtps2pdXrm|cvtss2sdXrm|cvtpd2psXrm|cvtsd2ssXrm", +"cvtdq2psXrm|cvttps2dqXrm|cvtps2dqXrm", +"subpsXrm|subssXrm|subpdXrm|subsdXrm","minpsXrm|minssXrm|minpdXrm|minsdXrm", +"divpsXrm|divssXrm|divpdXrm|divsdXrm","maxpsXrm|maxssXrm|maxpdXrm|maxsdXrm", +--6x +"punpcklbwPrm","punpcklwdPrm","punpckldqPrm","packsswbPrm", +"pcmpgtbPrm","pcmpgtwPrm","pcmpgtdPrm","packuswbPrm", +"punpckhbwPrm","punpckhwdPrm","punpckhdqPrm","packssdwPrm", +"||punpcklqdqXrm","||punpckhqdqXrm", +"movPrVSm","movqMrm|movdquXrm|movdqaXrm", +--7x +"pshufwMrmu|pshufhwXrmu|pshufdXrmu|pshuflwXrmu","pshiftw!Pmu", +"pshiftd!Pmu","pshiftq!Mmu||pshiftdq!Xmu", +"pcmpeqbPrm","pcmpeqwPrm","pcmpeqdPrm","emms|", +"vmreadUmr||extrqXmuu$|insertqXrmuu$","vmwriteUrm||extrqXrm$|insertqXrm$", +nil,nil, +"||haddpdXrm|haddpsXrm","||hsubpdXrm|hsubpsXrm", +"movVSmMr|movqXrm|movVSmXr","movqMmr|movdquXmr|movdqaXmr", +--8x +"joVj","jnoVj","jbVj","jnbVj","jzVj","jnzVj","jbeVj","jaVj", +"jsVj","jnsVj","jpeVj","jpoVj","jlVj","jgeVj","jleVj","jgVj", +--9x +"setoBm","setnoBm","setbBm","setnbBm","setzBm","setnzBm","setbeBm","setaBm", +"setsBm","setnsBm","setpeBm","setpoBm","setlBm","setgeBm","setleBm","setgBm", +--Ax +"push fs","pop fs","cpuid","btVmr","shldVmru","shldVmrc",nil,nil, +"push gs","pop gs","rsm","btsVmr","shrdVmru","shrdVmrc","fxsave!Dmp","imulVrm", +--Bx +"cmpxchgBmr","cmpxchgVmr","$lssVrm","btrVmr", +"$lfsVrm","$lgsVrm","movzxVrBmt","movzxVrWmt", +"|popcntVrm","ud2Dp","bt!Vmu","btcVmr", +"bsfVrm","bsrVrm|lzcntVrm|bsrWrm","movsxVrBmt","movsxVrWmt", +--Cx +"xaddBmr","xaddVmr", +"cmppsXrmu|cmpssXrmu|cmppdXrmu|cmpsdXrmu","$movntiVmr|", +"pinsrwPrWmu","pextrwDrPmu", +"shufpsXrmu||shufpdXrmu","$cmpxchg!Qmp", +"bswapVR","bswapVR","bswapVR","bswapVR","bswapVR","bswapVR","bswapVR","bswapVR", +--Dx +"||addsubpdXrm|addsubpsXrm","psrlwPrm","psrldPrm","psrlqPrm", +"paddqPrm","pmullwPrm", +"|movq2dqXrMm|movqXmr|movdq2qMrXm$","pmovmskbVrMm||pmovmskbVrXm", +"psubusbPrm","psubuswPrm","pminubPrm","pandPrm", +"paddusbPrm","padduswPrm","pmaxubPrm","pandnPrm", +--Ex +"pavgbPrm","psrawPrm","psradPrm","pavgwPrm", +"pmulhuwPrm","pmulhwPrm", +"|cvtdq2pdXrm|cvttpd2dqXrm|cvtpd2dqXrm","$movntqMmr||$movntdqXmr", +"psubsbPrm","psubswPrm","pminswPrm","porPrm", +"paddsbPrm","paddswPrm","pmaxswPrm","pxorPrm", +--Fx +"|||lddquXrm","psllwPrm","pslldPrm","psllqPrm", +"pmuludqPrm","pmaddwdPrm","psadbwPrm","maskmovqMrm||maskmovdquXrm$", +"psubbPrm","psubwPrm","psubdPrm","psubqPrm", +"paddbPrm","paddwPrm","padddPrm","ud", +} +assert(map_opc2[255] == "ud") + +-- Map for three-byte opcodes. Can't wait for their next invention. +local map_opc3 = { +["38"] = { -- [66] 0f 38 xx +--0x +[0]="pshufbPrm","phaddwPrm","phadddPrm","phaddswPrm", +"pmaddubswPrm","phsubwPrm","phsubdPrm","phsubswPrm", +"psignbPrm","psignwPrm","psigndPrm","pmulhrswPrm", +nil,nil,nil,nil, +--1x +"||pblendvbXrma",nil,nil,nil, +"||blendvpsXrma","||blendvpdXrma",nil,"||ptestXrm", +nil,nil,nil,nil, +"pabsbPrm","pabswPrm","pabsdPrm",nil, +--2x +"||pmovsxbwXrm","||pmovsxbdXrm","||pmovsxbqXrm","||pmovsxwdXrm", +"||pmovsxwqXrm","||pmovsxdqXrm",nil,nil, +"||pmuldqXrm","||pcmpeqqXrm","||$movntdqaXrm","||packusdwXrm", +nil,nil,nil,nil, +--3x +"||pmovzxbwXrm","||pmovzxbdXrm","||pmovzxbqXrm","||pmovzxwdXrm", +"||pmovzxwqXrm","||pmovzxdqXrm",nil,"||pcmpgtqXrm", +"||pminsbXrm","||pminsdXrm","||pminuwXrm","||pminudXrm", +"||pmaxsbXrm","||pmaxsdXrm","||pmaxuwXrm","||pmaxudXrm", +--4x +"||pmulddXrm","||phminposuwXrm", +--Fx +[0xf0] = "|||crc32TrBmt",[0xf1] = "|||crc32TrVmt", +}, + +["3a"] = { -- [66] 0f 3a xx +--0x +[0x00]=nil,nil,nil,nil,nil,nil,nil,nil, +"||roundpsXrmu","||roundpdXrmu","||roundssXrmu","||roundsdXrmu", +"||blendpsXrmu","||blendpdXrmu","||pblendwXrmu","palignrPrmu", +--1x +nil,nil,nil,nil, +"||pextrbVmXru","||pextrwVmXru","||pextrVmSXru","||extractpsVmXru", +nil,nil,nil,nil,nil,nil,nil,nil, +--2x +"||pinsrbXrVmu","||insertpsXrmu","||pinsrXrVmuS",nil, +--4x +[0x40] = "||dppsXrmu", +[0x41] = "||dppdXrmu", +[0x42] = "||mpsadbwXrmu", +--6x +[0x60] = "||pcmpestrmXrmu",[0x61] = "||pcmpestriXrmu", +[0x62] = "||pcmpistrmXrmu",[0x63] = "||pcmpistriXrmu", +}, +} + +-- Map for VMX/SVM opcodes 0F 01 C0-FF (sgdt group with register operands). +local map_opcvm = { +[0xc1]="vmcall",[0xc2]="vmlaunch",[0xc3]="vmresume",[0xc4]="vmxoff", +[0xc8]="monitor",[0xc9]="mwait", +[0xd8]="vmrun",[0xd9]="vmmcall",[0xda]="vmload",[0xdb]="vmsave", +[0xdc]="stgi",[0xdd]="clgi",[0xde]="skinit",[0xdf]="invlpga", +[0xf8]="swapgs",[0xf9]="rdtscp", +} + +-- Map for FP opcodes. And you thought stack machines are simple? +local map_opcfp = { +-- D8-DF 00-BF: opcodes with a memory operand. +-- D8 +[0]="faddFm","fmulFm","fcomFm","fcompFm","fsubFm","fsubrFm","fdivFm","fdivrFm", +"fldFm",nil,"fstFm","fstpFm","fldenvVm","fldcwWm","fnstenvVm","fnstcwWm", +-- DA +"fiaddDm","fimulDm","ficomDm","ficompDm", +"fisubDm","fisubrDm","fidivDm","fidivrDm", +-- DB +"fildDm","fisttpDm","fistDm","fistpDm",nil,"fld twordFmp",nil,"fstp twordFmp", +-- DC +"faddGm","fmulGm","fcomGm","fcompGm","fsubGm","fsubrGm","fdivGm","fdivrGm", +-- DD +"fldGm","fisttpQm","fstGm","fstpGm","frstorDmp",nil,"fnsaveDmp","fnstswWm", +-- DE +"fiaddWm","fimulWm","ficomWm","ficompWm", +"fisubWm","fisubrWm","fidivWm","fidivrWm", +-- DF +"fildWm","fisttpWm","fistWm","fistpWm", +"fbld twordFmp","fildQm","fbstp twordFmp","fistpQm", +-- xx C0-FF: opcodes with a pseudo-register operand. +-- D8 +"faddFf","fmulFf","fcomFf","fcompFf","fsubFf","fsubrFf","fdivFf","fdivrFf", +-- D9 +"fldFf","fxchFf",{"fnop"},nil, +{"fchs","fabs",nil,nil,"ftst","fxam"}, +{"fld1","fldl2t","fldl2e","fldpi","fldlg2","fldln2","fldz"}, +{"f2xm1","fyl2x","fptan","fpatan","fxtract","fprem1","fdecstp","fincstp"}, +{"fprem","fyl2xp1","fsqrt","fsincos","frndint","fscale","fsin","fcos"}, +-- DA +"fcmovbFf","fcmoveFf","fcmovbeFf","fcmovuFf",nil,{nil,"fucompp"},nil,nil, +-- DB +"fcmovnbFf","fcmovneFf","fcmovnbeFf","fcmovnuFf", +{nil,nil,"fnclex","fninit"},"fucomiFf","fcomiFf",nil, +-- DC +"fadd toFf","fmul toFf",nil,nil, +"fsub toFf","fsubr toFf","fdivr toFf","fdiv toFf", +-- DD +"ffreeFf",nil,"fstFf","fstpFf","fucomFf","fucompFf",nil,nil, +-- DE +"faddpFf","fmulpFf",nil,{nil,"fcompp"}, +"fsubrpFf","fsubpFf","fdivrpFf","fdivpFf", +-- DF +nil,nil,nil,nil,{"fnstsw ax"},"fucomipFf","fcomipFf",nil, +} +assert(map_opcfp[126] == "fcomipFf") + +-- Map for opcode groups. The subkey is sp from the ModRM byte. +local map_opcgroup = { + arith = { "add", "or", "adc", "sbb", "and", "sub", "xor", "cmp" }, + shift = { "rol", "ror", "rcl", "rcr", "shl", "shr", "sal", "sar" }, + testb = { "testBmi", "testBmi", "not", "neg", "mul", "imul", "div", "idiv" }, + testv = { "testVmi", "testVmi", "not", "neg", "mul", "imul", "div", "idiv" }, + incb = { "inc", "dec" }, + incd = { "inc", "dec", "callUmp", "$call farDmp", + "jmpUmp", "$jmp farDmp", "pushUm" }, + sldt = { "sldt", "str", "lldt", "ltr", "verr", "verw" }, + sgdt = { "vm*$sgdt", "vm*$sidt", "$lgdt", "vm*$lidt", + "smsw", nil, "lmsw", "vm*$invlpg" }, + bt = { nil, nil, nil, nil, "bt", "bts", "btr", "btc" }, + cmpxchg = { nil, "sz*,cmpxchg8bQmp,cmpxchg16bXmp", nil, nil, + nil, nil, "vmptrld|vmxon|vmclear", "vmptrst" }, + pshiftw = { nil, nil, "psrlw", nil, "psraw", nil, "psllw" }, + pshiftd = { nil, nil, "psrld", nil, "psrad", nil, "pslld" }, + pshiftq = { nil, nil, "psrlq", nil, nil, nil, "psllq" }, + pshiftdq = { nil, nil, "psrlq", "psrldq", nil, nil, "psllq", "pslldq" }, + fxsave = { "$fxsave", "$fxrstor", "$ldmxcsr", "$stmxcsr", + nil, "lfenceDp$", "mfenceDp$", "sfenceDp$clflush" }, + prefetch = { "prefetch", "prefetchw" }, + prefetcht = { "prefetchnta", "prefetcht0", "prefetcht1", "prefetcht2" }, +} + +------------------------------------------------------------------------------ + +-- Maps for register names. +local map_regs = { + B = { "al", "cl", "dl", "bl", "ah", "ch", "dh", "bh", + "r8b", "r9b", "r10b", "r11b", "r12b", "r13b", "r14b", "r15b" }, + B64 = { "al", "cl", "dl", "bl", "spl", "bpl", "sil", "dil", + "r8b", "r9b", "r10b", "r11b", "r12b", "r13b", "r14b", "r15b" }, + W = { "ax", "cx", "dx", "bx", "sp", "bp", "si", "di", + "r8w", "r9w", "r10w", "r11w", "r12w", "r13w", "r14w", "r15w" }, + D = { "eax", "ecx", "edx", "ebx", "esp", "ebp", "esi", "edi", + "r8d", "r9d", "r10d", "r11d", "r12d", "r13d", "r14d", "r15d" }, + Q = { "rax", "rcx", "rdx", "rbx", "rsp", "rbp", "rsi", "rdi", + "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15" }, + M = { "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7", + "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7" }, -- No x64 ext! + X = { "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7", + "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15" }, +} +local map_segregs = { "es", "cs", "ss", "ds", "fs", "gs", "segr6", "segr7" } + +-- Maps for size names. +local map_sz2n = { + B = 1, W = 2, D = 4, Q = 8, M = 8, X = 16, +} +local map_sz2prefix = { + B = "byte", W = "word", D = "dword", + Q = "qword", + M = "qword", X = "xword", + F = "dword", G = "qword", -- No need for sizes/register names for these two. +} + +------------------------------------------------------------------------------ + +-- Output a nicely formatted line with an opcode and operands. +local function putop(ctx, text, operands) + local code, pos, hex = ctx.code, ctx.pos, "" + local hmax = ctx.hexdump + if hmax > 0 then + for i=ctx.start,pos-1 do + hex = hex..format("%02X", byte(code, i, i)) + end + if #hex > hmax then hex = sub(hex, 1, hmax)..". " + else hex = hex..rep(" ", hmax-#hex+2) end + end + if operands then text = text.." "..operands end + if ctx.o16 then text = "o16 "..text; ctx.o16 = false end + if ctx.a32 then text = "a32 "..text; ctx.a32 = false end + if ctx.rep then text = ctx.rep.." "..text; ctx.rep = false end + if ctx.rex then + local t = (ctx.rexw and "w" or "")..(ctx.rexr and "r" or "").. + (ctx.rexx and "x" or "")..(ctx.rexb and "b" or "") + if t ~= "" then text = "rex."..t.." "..text end + ctx.rexw = false; ctx.rexr = false; ctx.rexx = false; ctx.rexb = false + ctx.rex = false + end + if ctx.seg then + local text2, n = gsub(text, "%[", "["..ctx.seg..":") + if n == 0 then text = ctx.seg.." "..text else text = text2 end + ctx.seg = false + end + if ctx.lock then text = "lock "..text; ctx.lock = false end + local imm = ctx.imm + if imm then + local sym = ctx.symtab[imm] + if sym then text = text.."\t->"..sym end + end + ctx.out(format("%08x %s%s\n", ctx.addr+ctx.start, hex, text)) + ctx.mrm = false + ctx.start = pos + ctx.imm = nil +end + +-- Clear all prefix flags. +local function clearprefixes(ctx) + ctx.o16 = false; ctx.seg = false; ctx.lock = false; ctx.rep = false + ctx.rexw = false; ctx.rexr = false; ctx.rexx = false; ctx.rexb = false + ctx.rex = false; ctx.a32 = false +end + +-- Fallback for incomplete opcodes at the end. +local function incomplete(ctx) + ctx.pos = ctx.stop+1 + clearprefixes(ctx) + return putop(ctx, "(incomplete)") +end + +-- Fallback for unknown opcodes. +local function unknown(ctx) + clearprefixes(ctx) + return putop(ctx, "(unknown)") +end + +-- Return an immediate of the specified size. +local function getimm(ctx, pos, n) + if pos+n-1 > ctx.stop then return incomplete(ctx) end + local code = ctx.code + if n == 1 then + local b1 = byte(code, pos, pos) + return b1 + elseif n == 2 then + local b1, b2 = byte(code, pos, pos+1) + return b1+b2*256 + else + local b1, b2, b3, b4 = byte(code, pos, pos+3) + local imm = b1+b2*256+b3*65536+b4*16777216 + ctx.imm = imm + return imm + end +end + +-- Process pattern string and generate the operands. +local function putpat(ctx, name, pat) + local operands, regs, sz, mode, sp, rm, sc, rx, sdisp + local code, pos, stop = ctx.code, ctx.pos, ctx.stop + + -- Chars used: 1DFGIMPQRSTUVWXacdfgijmoprstuwxyz + for p in gmatch(pat, ".") do + local x = nil + if p == "V" or p == "U" then + if ctx.rexw then sz = "Q"; ctx.rexw = false + elseif ctx.o16 then sz = "W"; ctx.o16 = false + elseif p == "U" and ctx.x64 then sz = "Q" + else sz = "D" end + regs = map_regs[sz] + elseif p == "T" then + if ctx.rexw then sz = "Q"; ctx.rexw = false else sz = "D" end + regs = map_regs[sz] + elseif p == "B" then + sz = "B" + regs = ctx.rex and map_regs.B64 or map_regs.B + elseif match(p, "[WDQMXFG]") then + sz = p + regs = map_regs[sz] + elseif p == "P" then + sz = ctx.o16 and "X" or "M"; ctx.o16 = false + regs = map_regs[sz] + elseif p == "S" then + name = name..lower(sz) + elseif p == "s" then + local imm = getimm(ctx, pos, 1); if not imm then return end + x = imm <= 127 and format("+0x%02x", imm) + or format("-0x%02x", 256-imm) + pos = pos+1 + elseif p == "u" then + local imm = getimm(ctx, pos, 1); if not imm then return end + x = format("0x%02x", imm) + pos = pos+1 + elseif p == "w" then + local imm = getimm(ctx, pos, 2); if not imm then return end + x = format("0x%x", imm) + pos = pos+2 + elseif p == "o" then -- [offset] + if ctx.x64 then + local imm1 = getimm(ctx, pos, 4); if not imm1 then return end + local imm2 = getimm(ctx, pos+4, 4); if not imm2 then return end + x = format("[0x%08x%08x]", imm2, imm1) + pos = pos+8 + else + local imm = getimm(ctx, pos, 4); if not imm then return end + x = format("[0x%08x]", imm) + pos = pos+4 + end + elseif p == "i" or p == "I" then + local n = map_sz2n[sz] + if n == 8 and ctx.x64 and p == "I" then + local imm1 = getimm(ctx, pos, 4); if not imm1 then return end + local imm2 = getimm(ctx, pos+4, 4); if not imm2 then return end + x = format("0x%08x%08x", imm2, imm1) + else + if n == 8 then n = 4 end + local imm = getimm(ctx, pos, n); if not imm then return end + if sz == "Q" and (imm < 0 or imm > 0x7fffffff) then + imm = (0xffffffff+1)-imm + x = format(imm > 65535 and "-0x%08x" or "-0x%x", imm) + else + x = format(imm > 65535 and "0x%08x" or "0x%x", imm) + end + end + pos = pos+n + elseif p == "j" then + local n = map_sz2n[sz] + if n == 8 then n = 4 end + local imm = getimm(ctx, pos, n); if not imm then return end + if sz == "B" and imm > 127 then imm = imm-256 + elseif imm > 2147483647 then imm = imm-4294967296 end + pos = pos+n + imm = imm + pos + ctx.addr + if imm > 4294967295 and not ctx.x64 then imm = imm-4294967296 end + ctx.imm = imm + if sz == "W" then + x = format("word 0x%04x", imm%65536) + elseif ctx.x64 then + local lo = imm % 0x1000000 + x = format("0x%02x%06x", (imm-lo) / 0x1000000, lo) + else + x = format("0x%08x", imm) + end + elseif p == "R" then + local r = byte(code, pos-1, pos-1)%8 + if ctx.rexb then r = r + 8; ctx.rexb = false end + x = regs[r+1] + elseif p == "a" then x = regs[1] + elseif p == "c" then x = "cl" + elseif p == "d" then x = "dx" + elseif p == "1" then x = "1" + else + if not mode then + mode = ctx.mrm + if not mode then + if pos > stop then return incomplete(ctx) end + mode = byte(code, pos, pos) + pos = pos+1 + end + rm = mode%8; mode = (mode-rm)/8 + sp = mode%8; mode = (mode-sp)/8 + sdisp = "" + if mode < 3 then + if rm == 4 then + if pos > stop then return incomplete(ctx) end + sc = byte(code, pos, pos) + pos = pos+1 + rm = sc%8; sc = (sc-rm)/8 + rx = sc%8; sc = (sc-rx)/8 + if ctx.rexx then rx = rx + 8; ctx.rexx = false end + if rx == 4 then rx = nil end + end + if mode > 0 or rm == 5 then + local dsz = mode + if dsz ~= 1 then dsz = 4 end + local disp = getimm(ctx, pos, dsz); if not disp then return end + if mode == 0 then rm = nil end + if rm or rx or (not sc and ctx.x64 and not ctx.a32) then + if dsz == 1 and disp > 127 then + sdisp = format("-0x%x", 256-disp) + elseif disp >= 0 and disp <= 0x7fffffff then + sdisp = format("+0x%x", disp) + else + sdisp = format("-0x%x", (0xffffffff+1)-disp) + end + else + sdisp = format(ctx.x64 and not ctx.a32 and + not (disp >= 0 and disp <= 0x7fffffff) + and "0xffffffff%08x" or "0x%08x", disp) + end + pos = pos+dsz + end + end + if rm and ctx.rexb then rm = rm + 8; ctx.rexb = false end + if ctx.rexr then sp = sp + 8; ctx.rexr = false end + end + if p == "m" then + if mode == 3 then x = regs[rm+1] + else + local aregs = ctx.a32 and map_regs.D or ctx.aregs + local srm, srx = "", "" + if rm then srm = aregs[rm+1] + elseif not sc and ctx.x64 and not ctx.a32 then srm = "rip" end + ctx.a32 = false + if rx then + if rm then srm = srm.."+" end + srx = aregs[rx+1] + if sc > 0 then srx = srx.."*"..(2^sc) end + end + x = format("[%s%s%s]", srm, srx, sdisp) + end + if mode < 3 and + (not match(pat, "[aRrgp]") or match(pat, "t")) then -- Yuck. + x = map_sz2prefix[sz].." "..x + end + elseif p == "r" then x = regs[sp+1] + elseif p == "g" then x = map_segregs[sp+1] + elseif p == "p" then -- Suppress prefix. + elseif p == "f" then x = "st"..rm + elseif p == "x" then + if sp == 0 and ctx.lock and not ctx.x64 then + x = "CR8"; ctx.lock = false + else + x = "CR"..sp + end + elseif p == "y" then x = "DR"..sp + elseif p == "z" then x = "TR"..sp + elseif p == "t" then + else + error("bad pattern `"..pat.."'") + end + end + if x then operands = operands and operands..", "..x or x end + end + ctx.pos = pos + return putop(ctx, name, operands) +end + +-- Forward declaration. +local map_act + +-- Fetch and cache MRM byte. +local function getmrm(ctx) + local mrm = ctx.mrm + if not mrm then + local pos = ctx.pos + if pos > ctx.stop then return nil end + mrm = byte(ctx.code, pos, pos) + ctx.pos = pos+1 + ctx.mrm = mrm + end + return mrm +end + +-- Dispatch to handler depending on pattern. +local function dispatch(ctx, opat, patgrp) + if not opat then return unknown(ctx) end + if match(opat, "%|") then -- MMX/SSE variants depending on prefix. + local p + if ctx.rep then + p = ctx.rep=="rep" and "%|([^%|]*)" or "%|[^%|]*%|[^%|]*%|([^%|]*)" + ctx.rep = false + elseif ctx.o16 then p = "%|[^%|]*%|([^%|]*)"; ctx.o16 = false + else p = "^[^%|]*" end + opat = match(opat, p) + if not opat then return unknown(ctx) end +-- ctx.rep = false; ctx.o16 = false + --XXX fails for 66 f2 0f 38 f1 06 crc32 eax,WORD PTR [esi] + --XXX remove in branches? + end + if match(opat, "%$") then -- reg$mem variants. + local mrm = getmrm(ctx); if not mrm then return incomplete(ctx) end + opat = match(opat, mrm >= 192 and "^[^%$]*" or "%$(.*)") + if opat == "" then return unknown(ctx) end + end + if opat == "" then return unknown(ctx) end + local name, pat = match(opat, "^([a-z0-9 ]*)(.*)") + if pat == "" and patgrp then pat = patgrp end + return map_act[sub(pat, 1, 1)](ctx, name, pat) +end + +-- Get a pattern from an opcode map and dispatch to handler. +local function dispatchmap(ctx, opcmap) + local pos = ctx.pos + local opat = opcmap[byte(ctx.code, pos, pos)] + pos = pos + 1 + ctx.pos = pos + return dispatch(ctx, opat) +end + +-- Map for action codes. The key is the first char after the name. +map_act = { + -- Simple opcodes without operands. + [""] = function(ctx, name, pat) + return putop(ctx, name) + end, + + -- Operand size chars fall right through. + B = putpat, W = putpat, D = putpat, Q = putpat, + V = putpat, U = putpat, T = putpat, + M = putpat, X = putpat, P = putpat, + F = putpat, G = putpat, + + -- Collect prefixes. + [":"] = function(ctx, name, pat) + ctx[pat == ":" and name or sub(pat, 2)] = name + if ctx.pos - ctx.start > 5 then return unknown(ctx) end -- Limit #prefixes. + end, + + -- Chain to special handler specified by name. + ["*"] = function(ctx, name, pat) + return map_act[name](ctx, name, sub(pat, 2)) + end, + + -- Use named subtable for opcode group. + ["!"] = function(ctx, name, pat) + local mrm = getmrm(ctx); if not mrm then return incomplete(ctx) end + return dispatch(ctx, map_opcgroup[name][((mrm-(mrm%8))/8)%8+1], sub(pat, 2)) + end, + + -- o16,o32[,o64] variants. + sz = function(ctx, name, pat) + if ctx.o16 then ctx.o16 = false + else + pat = match(pat, ",(.*)") + if ctx.rexw then + local p = match(pat, ",(.*)") + if p then pat = p; ctx.rexw = false end + end + end + pat = match(pat, "^[^,]*") + return dispatch(ctx, pat) + end, + + -- Two-byte opcode dispatch. + opc2 = function(ctx, name, pat) + return dispatchmap(ctx, map_opc2) + end, + + -- Three-byte opcode dispatch. + opc3 = function(ctx, name, pat) + return dispatchmap(ctx, map_opc3[pat]) + end, + + -- VMX/SVM dispatch. + vm = function(ctx, name, pat) + return dispatch(ctx, map_opcvm[ctx.mrm]) + end, + + -- Floating point opcode dispatch. + fp = function(ctx, name, pat) + local mrm = getmrm(ctx); if not mrm then return incomplete(ctx) end + local rm = mrm%8 + local idx = pat*8 + ((mrm-rm)/8)%8 + if mrm >= 192 then idx = idx + 64 end + local opat = map_opcfp[idx] + if type(opat) == "table" then opat = opat[rm+1] end + return dispatch(ctx, opat) + end, + + -- REX prefix. + rex = function(ctx, name, pat) + if ctx.rex then return unknown(ctx) end -- Only 1 REX prefix allowed. + for p in gmatch(pat, ".") do ctx["rex"..p] = true end + ctx.rex = true + end, + + -- Special case for nop with REX prefix. + nop = function(ctx, name, pat) + return dispatch(ctx, ctx.rex and pat or "nop") + end, +} + +------------------------------------------------------------------------------ + +-- Disassemble a block of code. +local function disass_block(ctx, ofs, len) + if not ofs then ofs = 0 end + local stop = len and ofs+len or #ctx.code + ofs = ofs + 1 + ctx.start = ofs + ctx.pos = ofs + ctx.stop = stop + ctx.imm = nil + ctx.mrm = false + clearprefixes(ctx) + while ctx.pos <= stop do dispatchmap(ctx, ctx.map1) end + if ctx.pos ~= ctx.start then incomplete(ctx) end +end + +-- Extended API: create a disassembler context. Then call ctx:disass(ofs, len). +local function create_(code, addr, out) + local ctx = {} + ctx.code = code + ctx.addr = (addr or 0) - 1 + ctx.out = out or io.write + ctx.symtab = {} + ctx.disass = disass_block + ctx.hexdump = 16 + ctx.x64 = false + ctx.map1 = map_opc1_32 + ctx.aregs = map_regs.D + return ctx +end + +local function create64_(code, addr, out) + local ctx = create_(code, addr, out) + ctx.x64 = true + ctx.map1 = map_opc1_64 + ctx.aregs = map_regs.Q + return ctx +end + +-- Simple API: disassemble code (a string) at address and output via out. +local function disass_(code, addr, out) + create_(code, addr, out):disass() +end + +local function disass64_(code, addr, out) + create64_(code, addr, out):disass() +end + +-- Return register name for RID. +local function regname_(r) + if r < 8 then return map_regs.D[r+1] end + return map_regs.X[r-7] +end + +local function regname64_(r) + if r < 16 then return map_regs.Q[r+1] end + return map_regs.X[r-15] +end + +-- Public module functions. +module(...) + +create = create_ +create64 = create64_ +disass = disass_ +disass64 = disass64_ +regname = regname_ +regname64 = regname64_ + diff --git a/polymer/eduke32/source/lunatic/dump.lua b/polymer/eduke32/source/lunatic/dump.lua new file mode 100644 index 000000000..70a59280e --- /dev/null +++ b/polymer/eduke32/source/lunatic/dump.lua @@ -0,0 +1,700 @@ +---------------------------------------------------------------------------- +-- LuaJIT compiler dump module. +-- +-- Copyright (C) 2005-2013 Mike Pall. All rights reserved. +-- Released under the MIT license. See Copyright Notice in luajit.h +---------------------------------------------------------------------------- +-- +-- This module can be used to debug the JIT compiler itself. It dumps the +-- code representations and structures used in various compiler stages. +-- +-- Example usage: +-- +-- luajit -jdump -e "local x=0; for i=1,1e6 do x=x+i end; print(x)" +-- luajit -jdump=im -e "for i=1,1000 do for j=1,1000 do end end" | less -R +-- luajit -jdump=is myapp.lua | less -R +-- luajit -jdump=-b myapp.lua +-- luajit -jdump=+aH,myapp.html myapp.lua +-- luajit -jdump=ixT,myapp.dump myapp.lua +-- +-- The first argument specifies the dump mode. The second argument gives +-- the output file name. Default output is to stdout, unless the environment +-- variable LUAJIT_DUMPFILE is set. The file is overwritten every time the +-- module is started. +-- +-- Different features can be turned on or off with the dump mode. If the +-- mode starts with a '+', the following features are added to the default +-- set of features; a '-' removes them. Otherwise the features are replaced. +-- +-- The following dump features are available (* marks the default): +-- +-- * t Print a line for each started, ended or aborted trace (see also -jv). +-- * b Dump the traced bytecode. +-- * i Dump the IR (intermediate representation). +-- r Augment the IR with register/stack slots. +-- s Dump the snapshot map. +-- * m Dump the generated machine code. +-- x Print each taken trace exit. +-- X Print each taken trace exit and the contents of all registers. +-- +-- The output format can be set with the following characters: +-- +-- T Plain text output. +-- A ANSI-colored text output +-- H Colorized HTML + CSS output. +-- +-- The default output format is plain text. It's set to ANSI-colored text +-- if the COLORTERM variable is set. Note: this is independent of any output +-- redirection, which is actually considered a feature. +-- +-- You probably want to use less -R to enjoy viewing ANSI-colored text from +-- a pipe or a file. Add this to your ~/.bashrc: export LESS="-R" +-- +------------------------------------------------------------------------------ + +-- Cache some library functions and objects. +local jit = require("jit") +assert(jit.version_num == 20001, "LuaJIT core/library version mismatch") +local jutil = require("jit.util") +local vmdef = require("jit.vmdef") +local funcinfo, funcbc = jutil.funcinfo, jutil.funcbc +local traceinfo, traceir, tracek = jutil.traceinfo, jutil.traceir, jutil.tracek +local tracemc, tracesnap = jutil.tracemc, jutil.tracesnap +local traceexitstub, ircalladdr = jutil.traceexitstub, jutil.ircalladdr +local bit = require("bit") +local band, shl, shr = bit.band, bit.lshift, bit.rshift +local sub, gsub, format = string.sub, string.gsub, string.format +local byte, char, rep = string.byte, string.char, string.rep +local type, tostring = type, tostring +local stdout, stderr = io.stdout, io.stderr + +-- Load other modules on-demand. +local bcline, disass + +-- Active flag, output file handle and dump mode. +local active, out, dumpmode + +------------------------------------------------------------------------------ + +local symtabmt = { __index = false } +local symtab = {} +local nexitsym = 0 + +-- Fill nested symbol table with per-trace exit stub addresses. +local function fillsymtab_tr(tr, nexit) + local t = {} + symtabmt.__index = t + if jit.arch == "mips" or jit.arch == "mipsel" then + t[traceexitstub(tr, 0)] = "exit" + return + end + for i=0,nexit-1 do + local addr = traceexitstub(tr, i) + t[addr] = tostring(i) + end + local addr = traceexitstub(tr, nexit) + if addr then t[addr] = "stack_check" end +end + +-- Fill symbol table with trace exit stub addresses. +local function fillsymtab(tr, nexit) + local t = symtab + if nexitsym == 0 then + local ircall = vmdef.ircall + for i=0,#ircall do + local addr = ircalladdr(i) + if addr ~= 0 then t[addr] = ircall[i] end + end + end + if nexitsym == 1000000 then -- Per-trace exit stubs. + fillsymtab_tr(tr, nexit) + elseif nexit > nexitsym then -- Shared exit stubs. + for i=nexitsym,nexit-1 do + local addr = traceexitstub(i) + if addr == nil then -- Fall back to per-trace exit stubs. + fillsymtab_tr(tr, nexit) + setmetatable(symtab, symtabmt) + nexit = 1000000 + break + end + t[addr] = tostring(i) + end + nexitsym = nexit + end + return t +end + +local function dumpwrite(s) + out:write(s) +end + +-- Disassemble machine code. +local function dump_mcode(tr) + local info = traceinfo(tr) + if not info then return end + local mcode, addr, loop = tracemc(tr) + if not mcode then return end + if not disass then disass = require("jit.dis_"..jit.arch) end + out:write("---- TRACE ", tr, " mcode ", #mcode, "\n") + local ctx = disass.create(mcode, addr, dumpwrite) + ctx.hexdump = 0 + ctx.symtab = fillsymtab(tr, info.nexit) + if loop ~= 0 then + symtab[addr+loop] = "LOOP" + ctx:disass(0, loop) + out:write("->LOOP:\n") + ctx:disass(loop, #mcode-loop) + symtab[addr+loop] = nil + else + ctx:disass(0, #mcode) + end +end + +------------------------------------------------------------------------------ + +local irtype_text = { + [0] = "nil", + "fal", + "tru", + "lud", + "str", + "p32", + "thr", + "pro", + "fun", + "p64", + "cdt", + "tab", + "udt", + "flt", + "num", + "i8 ", + "u8 ", + "i16", + "u16", + "int", + "u32", + "i64", + "u64", + "sfp", +} + +local colortype_ansi = { + [0] = "%s", + "%s", + "%s", + "\027[36m%s\027[m", + "\027[32m%s\027[m", + "%s", + "\027[1m%s\027[m", + "%s", + "\027[1m%s\027[m", + "%s", + "\027[33m%s\027[m", + "\027[31m%s\027[m", + "\027[36m%s\027[m", + "\027[34m%s\027[m", + "\027[34m%s\027[m", + "\027[35m%s\027[m", + "\027[35m%s\027[m", + "\027[35m%s\027[m", + "\027[35m%s\027[m", + "\027[35m%s\027[m", + "\027[35m%s\027[m", + "\027[35m%s\027[m", + "\027[35m%s\027[m", + "\027[35m%s\027[m", +} + +local function colorize_text(s, t) + return s +end + +local function colorize_ansi(s, t) + return format(colortype_ansi[t], s) +end + +local irtype_ansi = setmetatable({}, + { __index = function(tab, t) + local s = colorize_ansi(irtype_text[t], t); tab[t] = s; return s; end }) + +local html_escape = { ["<"] = "<", [">"] = ">", ["&"] = "&", } + +local function colorize_html(s, t) + s = gsub(s, "[<>&]", html_escape) + return format('%s', irtype_text[t], s) +end + +local irtype_html = setmetatable({}, + { __index = function(tab, t) + local s = colorize_html(irtype_text[t], t); tab[t] = s; return s; end }) + +local header_html = [[ + +]] + +local colorize, irtype + +-- Lookup tables to convert some literals into names. +local litname = { + ["SLOAD "] = setmetatable({}, { __index = function(t, mode) + local s = "" + if band(mode, 1) ~= 0 then s = s.."P" end + if band(mode, 2) ~= 0 then s = s.."F" end + if band(mode, 4) ~= 0 then s = s.."T" end + if band(mode, 8) ~= 0 then s = s.."C" end + if band(mode, 16) ~= 0 then s = s.."R" end + if band(mode, 32) ~= 0 then s = s.."I" end + t[mode] = s + return s + end}), + ["XLOAD "] = { [0] = "", "R", "V", "RV", "U", "RU", "VU", "RVU", }, + ["CONV "] = setmetatable({}, { __index = function(t, mode) + local s = irtype[band(mode, 31)] + s = irtype[band(shr(mode, 5), 31)].."."..s + if band(mode, 0x400) ~= 0 then s = s.." trunc" + elseif band(mode, 0x800) ~= 0 then s = s.." sext" end + local c = shr(mode, 14) + if c == 2 then s = s.." index" elseif c == 3 then s = s.." check" end + t[mode] = s + return s + end}), + ["FLOAD "] = vmdef.irfield, + ["FREF "] = vmdef.irfield, + ["FPMATH"] = vmdef.irfpm, +} + +local function ctlsub(c) + if c == "\n" then return "\\n" + elseif c == "\r" then return "\\r" + elseif c == "\t" then return "\\t" + else return format("\\%03d", byte(c)) + end +end + +local function fmtfunc(func, pc) + local fi = funcinfo(func, pc) + if fi.loc then + return fi.loc + elseif fi.ffid then + return vmdef.ffnames[fi.ffid] + elseif fi.addr then + return format("C:%x", fi.addr) + else + return "(?)" + end +end + +local function formatk(tr, idx) + local k, t, slot = tracek(tr, idx) + local tn = type(k) + local s + if tn == "number" then + if k == 2^52+2^51 then + s = "bias" + else + s = format("%+.14g", k) + end + elseif tn == "string" then + s = format(#k > 20 and '"%.20s"~' or '"%s"', gsub(k, "%c", ctlsub)) + elseif tn == "function" then + s = fmtfunc(k) + elseif tn == "table" then + s = format("{%p}", k) + elseif tn == "userdata" then + if t == 12 then + s = format("userdata:%p", k) + else + s = format("[%p]", k) + if s == "[0x00000000]" then s = "NULL" end + end + elseif t == 21 then -- int64_t + s = sub(tostring(k), 1, -3) + if sub(s, 1, 1) ~= "-" then s = "+"..s end + else + s = tostring(k) -- For primitives. + end + s = colorize(format("%-4s", s), t) + if slot then + s = format("%s @%d", s, slot) + end + return s +end + +local function printsnap(tr, snap) + local n = 2 + for s=0,snap[1]-1 do + local sn = snap[n] + if shr(sn, 24) == s then + n = n + 1 + local ref = band(sn, 0xffff) - 0x8000 -- REF_BIAS + if ref < 0 then + out:write(formatk(tr, ref)) + elseif band(sn, 0x80000) ~= 0 then -- SNAP_SOFTFPNUM + out:write(colorize(format("%04d/%04d", ref, ref+1), 14)) + else + local m, ot, op1, op2 = traceir(tr, ref) + out:write(colorize(format("%04d", ref), band(ot, 31))) + end + out:write(band(sn, 0x10000) == 0 and " " or "|") -- SNAP_FRAME + else + out:write("---- ") + end + end + out:write("]\n") +end + +-- Dump snapshots (not interleaved with IR). +local function dump_snap(tr) + out:write("---- TRACE ", tr, " snapshots\n") + for i=0,1000000000 do + local snap = tracesnap(tr, i) + if not snap then break end + out:write(format("#%-3d %04d [ ", i, snap[0])) + printsnap(tr, snap) + end +end + +-- Return a register name or stack slot for a rid/sp location. +local function ridsp_name(ridsp, ins) + if not disass then disass = require("jit.dis_"..jit.arch) end + local rid, slot = band(ridsp, 0xff), shr(ridsp, 8) + if rid == 253 or rid == 254 then + return (slot == 0 or slot == 255) and " {sink" or format(" {%04d", ins-slot) + end + if ridsp > 255 then return format("[%x]", slot*4) end + if rid < 128 then return disass.regname(rid) end + return "" +end + +-- Dump CALL* function ref and return optional ctype. +local function dumpcallfunc(tr, ins) + local ctype + if ins > 0 then + local m, ot, op1, op2 = traceir(tr, ins) + if band(ot, 31) == 0 then -- nil type means CARG(func, ctype). + ins = op1 + ctype = formatk(tr, op2) + end + end + if ins < 0 then + out:write(format("[0x%x](", tonumber((tracek(tr, ins))))) + else + out:write(format("%04d (", ins)) + end + return ctype +end + +-- Recursively gather CALL* args and dump them. +local function dumpcallargs(tr, ins) + if ins < 0 then + out:write(formatk(tr, ins)) + else + local m, ot, op1, op2 = traceir(tr, ins) + local oidx = 6*shr(ot, 8) + local op = sub(vmdef.irnames, oidx+1, oidx+6) + if op == "CARG " then + dumpcallargs(tr, op1) + if op2 < 0 then + out:write(" ", formatk(tr, op2)) + else + out:write(" ", format("%04d", op2)) + end + else + out:write(format("%04d", ins)) + end + end +end + +-- Dump IR and interleaved snapshots. +local function dump_ir(tr, dumpsnap, dumpreg) + local info = traceinfo(tr) + if not info then return end + local nins = info.nins + out:write("---- TRACE ", tr, " IR\n") + local irnames = vmdef.irnames + local snapref = 65536 + local snap, snapno + if dumpsnap then + snap = tracesnap(tr, 0) + snapref = snap[0] + snapno = 0 + end + for ins=1,nins do + if ins >= snapref then + if dumpreg then + out:write(format(".... SNAP #%-3d [ ", snapno)) + else + out:write(format(".... SNAP #%-3d [ ", snapno)) + end + printsnap(tr, snap) + snapno = snapno + 1 + snap = tracesnap(tr, snapno) + snapref = snap and snap[0] or 65536 + end + local m, ot, op1, op2, ridsp = traceir(tr, ins) + local oidx, t = 6*shr(ot, 8), band(ot, 31) + local op = sub(irnames, oidx+1, oidx+6) + if op == "LOOP " then + if dumpreg then + out:write(format("%04d ------------ LOOP ------------\n", ins)) + else + out:write(format("%04d ------ LOOP ------------\n", ins)) + end + elseif op ~= "NOP " and op ~= "CARG " and + (dumpreg or op ~= "RENAME") then + local rid = band(ridsp, 255) + if dumpreg then + out:write(format("%04d %-6s", ins, ridsp_name(ridsp, ins))) + else + out:write(format("%04d ", ins)) + end + out:write(format("%s%s %s %s ", + (rid == 254 or rid == 253) and "}" or + (band(ot, 128) == 0 and " " or ">"), + band(ot, 64) == 0 and " " or "+", + irtype[t], op)) + local m1, m2 = band(m, 3), band(m, 3*4) + if sub(op, 1, 4) == "CALL" then + local ctype + if m2 == 1*4 then -- op2 == IRMlit + out:write(format("%-10s (", vmdef.ircall[op2])) + else + ctype = dumpcallfunc(tr, op2) + end + if op1 ~= -1 then dumpcallargs(tr, op1) end + out:write(")") + if ctype then out:write(" ctype ", ctype) end + elseif op == "CNEW " and op2 == -1 then + out:write(formatk(tr, op1)) + elseif m1 ~= 3 then -- op1 != IRMnone + if op1 < 0 then + out:write(formatk(tr, op1)) + else + out:write(format(m1 == 0 and "%04d" or "#%-3d", op1)) + end + if m2 ~= 3*4 then -- op2 != IRMnone + if m2 == 1*4 then -- op2 == IRMlit + local litn = litname[op] + if litn and litn[op2] then + out:write(" ", litn[op2]) + elseif op == "UREFO " or op == "UREFC " then + out:write(format(" #%-3d", shr(op2, 8))) + else + out:write(format(" #%-3d", op2)) + end + elseif op2 < 0 then + out:write(" ", formatk(tr, op2)) + else + out:write(format(" %04d", op2)) + end + end + end + out:write("\n") + end + end + if snap then + if dumpreg then + out:write(format(".... SNAP #%-3d [ ", snapno)) + else + out:write(format(".... SNAP #%-3d [ ", snapno)) + end + printsnap(tr, snap) + end +end + +------------------------------------------------------------------------------ + +local recprefix = "" +local recdepth = 0 + +-- Format trace error message. +local function fmterr(err, info) + if type(err) == "number" then + if type(info) == "function" then info = fmtfunc(info) end + err = format(vmdef.traceerr[err], info) + end + return err +end + +-- Dump trace states. +local function dump_trace(what, tr, func, pc, otr, oex) + if what == "stop" or (what == "abort" and dumpmode.a) then + if dumpmode.i then dump_ir(tr, dumpmode.s, dumpmode.r and what == "stop") + elseif dumpmode.s then dump_snap(tr) end + if dumpmode.m then dump_mcode(tr) end + end + if what == "start" then + if dumpmode.H then out:write('
\n') end
+    out:write("---- TRACE ", tr, " ", what)
+    if otr then out:write(" ", otr, "/", oex) end
+    out:write(" ", fmtfunc(func, pc), "\n")
+    recprefix = ""
+  elseif what == "stop" or what == "abort" then
+    out:write("---- TRACE ", tr, " ", what)
+    recprefix = nil
+    if what == "abort" then
+      out:write(" ", fmtfunc(func, pc), " -- ", fmterr(otr, oex), "\n")
+    else
+      local info = traceinfo(tr)
+      local link, ltype = info.link, info.linktype
+      if link == tr or link == 0 then
+	out:write(" -> ", ltype, "\n")
+      elseif ltype == "root" then
+	out:write(" -> ", link, "\n")
+      else
+	out:write(" -> ", link, " ", ltype, "\n")
+      end
+    end
+    if dumpmode.H then out:write("
\n\n") else out:write("\n") end + else + out:write("---- TRACE ", what, "\n\n") + end + out:flush() +end + +-- Dump recorded bytecode. +local function dump_record(tr, func, pc, depth, callee) + if depth ~= recdepth then + recdepth = depth + recprefix = rep(" .", depth) + end + local line + if pc >= 0 then + line = bcline(func, pc, recprefix) + if dumpmode.H then line = gsub(line, "[<>&]", html_escape) end + else + line = "0000 "..recprefix.." FUNCC \n" + callee = func + end + if pc <= 0 then + out:write(sub(line, 1, -2), " ; ", fmtfunc(func), "\n") + else + out:write(line) + end + if pc >= 0 and band(funcbc(func, pc), 0xff) < 16 then -- ORDER BC + out:write(bcline(func, pc+1, recprefix)) -- Write JMP for cond. + end +end + +------------------------------------------------------------------------------ + +-- Dump taken trace exits. +local function dump_texit(tr, ex, ngpr, nfpr, ...) + out:write("---- TRACE ", tr, " exit ", ex, "\n") + if dumpmode.X then + local regs = {...} + if jit.arch == "x64" then + for i=1,ngpr do + out:write(format(" %016x", regs[i])) + if i % 4 == 0 then out:write("\n") end + end + else + for i=1,ngpr do + out:write(format(" %08x", regs[i])) + if i % 8 == 0 then out:write("\n") end + end + end + if jit.arch == "mips" or jit.arch == "mipsel" then + for i=1,nfpr,2 do + out:write(format(" %+17.14g", regs[ngpr+i])) + if i % 8 == 7 then out:write("\n") end + end + else + for i=1,nfpr do + out:write(format(" %+17.14g", regs[ngpr+i])) + if i % 4 == 0 then out:write("\n") end + end + end + end +end + +------------------------------------------------------------------------------ + +-- Detach dump handlers. +local function dumpoff() + if active then + active = false + jit.attach(dump_texit) + jit.attach(dump_record) + jit.attach(dump_trace) + if out and out ~= stdout and out ~= stderr then out:close() end + out = nil + end +end + +-- Open the output file and attach dump handlers. +local function dumpon(opt, outfile) + if active then dumpoff() end + + local colormode = os.getenv("COLORTERM") and "A" or "T" + if opt then + opt = gsub(opt, "[TAH]", function(mode) colormode = mode; return ""; end) + end + + local m = { t=true, b=true, i=true, m=true, } + if opt and opt ~= "" then + local o = sub(opt, 1, 1) + if o ~= "+" and o ~= "-" then m = {} end + for i=1,#opt do m[sub(opt, i, i)] = (o ~= "-") end + end + dumpmode = m + + if m.t or m.b or m.i or m.s or m.m then + jit.attach(dump_trace, "trace") + end + if m.b then + jit.attach(dump_record, "record") + if not bcline then bcline = require("jit.bc").line end + end + if m.x or m.X then + jit.attach(dump_texit, "texit") + end + + if not outfile then outfile = os.getenv("LUAJIT_DUMPFILE") end + if outfile then + out = outfile == "-" and stdout or assert(io.open(outfile, "w")) + else + out = stdout + end + + m[colormode] = true + if colormode == "A" then + colorize = colorize_ansi + irtype = irtype_ansi + elseif colormode == "H" then + colorize = colorize_html + irtype = irtype_html + out:write(header_html) + else + colorize = colorize_text + irtype = irtype_text + end + + active = true +end + +-- Public module functions. +module(...) + +on = dumpon +off = dumpoff +start = dumpon -- For -j command line option. + diff --git a/polymer/eduke32/source/lunatic/dynsymlist b/polymer/eduke32/source/lunatic/dynsymlist index f82247254..4d09793e2 100644 --- a/polymer/eduke32/source/lunatic/dynsymlist +++ b/polymer/eduke32/source/lunatic/dynsymlist @@ -144,6 +144,9 @@ luaJIT_BC_bcheck; luaJIT_BC_xmath; luaJIT_BC_defs; luaJIT_BC_v; +luaJIT_BC_dump; +luaJIT_BC_dis_x86; +luaJIT_BC_dis_x64; rand_jkiss_u32; rand_jkiss_dbl;