|//
|// Common DynASM definitions and macros for x86 CPUs.
|// Copyright (C) 2005-2011 Mike Pall. See Copyright Notice in luajit.h
|//
|
|// Standard DynASM declarations.
|.arch x86
|.section code, deopt, tail, mfmap
|
|// Type definitions with (almost) global validity.
|.type L,		lua_State,	esi	// L.
|.type BASE,		TValue,		ebx	// L->base.
|.type TOP,		TValue,		edi	// L->top (calls/open ops).
|.type CI,		CallInfo,	ecx	// L->ci (calls, locally).
|.type LCL,		LClosure,	eax	// func->value (calls).
|
|// Type definitions with local validity.
|.type GL,		global_State
|.type TVALUE,		TValue
|.type VALUE,		Value
|.type CINFO,		CallInfo
|.type GCOBJECT,	GCObject
|.type TSTRING,		TString
|.type TABLE,		Table
|.type CCLOSURE,	CClosure
|.type PROTO,		Proto
|.type UPVAL,		UpVal
|.type NODE,		Node
|
|// Definitions copied to DynASM domain to avoid unnecessary constant args.
|// CHECK: must match with the definitions in lua.h!
|.define LUA_TNIL,		0
|.define LUA_TBOOLEAN,		1
|.define LUA_TLIGHTUSERDATA,	2
|.define LUA_TNUMBER,		3
|.define LUA_TSTRING,		4
|.define LUA_TTABLE,		5
|.define LUA_TFUNCTION,		6
|.define LUA_TUSERDATA,		7
|.define LUA_TTHREAD,		8
|
|.define LUA_TNUM_NUM,		0x33
|.define LUA_TNUM_NUM_NUM,	0x333
|.define LUA_TSTR_STR,		0x44
|.define LUA_TSTR_NUM,		0x43
|.define LUA_TSTR_NUM_NUM,	0x433
|.define LUA_TTABLE_NUM,	0x53
|.define LUA_TTABLE_STR,	0x54
|
|// Macros to test, set and copy stack slots.
|.macro istt, idx, tp;  cmp dword BASE[idx].tt, tp; .endmacro
|.macro isnil, idx;  istt idx, LUA_TNIL; .endmacro
|.macro isnumber, idx;  istt idx, LUA_TNUMBER; .endmacro
|.macro isstring, idx;  istt idx, LUA_TSTRING; .endmacro
|.macro istable, idx;  istt idx, LUA_TTABLE; .endmacro
|.macro isfunction, idx;  istt idx, LUA_TFUNCTION; .endmacro
|
|.macro isnumber2, idx1, idx2, reg
|  mov reg, BASE[idx1].tt;  shl reg, 4;  or reg, BASE[idx2].tt
|  cmp reg, LUA_TNUM_NUM
|.endmacro
|.macro isnumber2, idx1, idx2; isnumber2, idx1, idx2, eax; .endmacro
|
|.macro isnumber3, idx1, idx2, idx3, reg
|  mov reg, BASE[idx1].tt;  shl reg, 4;  or reg, BASE[idx2].tt
|  shl reg, 4;  or reg, BASE[idx3].tt;  cmp reg, LUA_TNUM_NUM_NUM
|.endmacro
|.macro isnumber3, idx1, idx2, idx3; isnumber3, idx1, idx2, idx3, eax; .endmacro
|
|.macro tvisnil, tv;  cmp dword tv.tt, LUA_TNIL; .endmacro
|
|.macro settt, tv, tp;  mov dword tv.tt, tp; .endmacro
|.macro setnilvalue, tv;  settt tv, LUA_TNIL; .endmacro
|
|.macro setbvalue, tv, val		// May use edx.
||if (val) {  /* true */
|   mov edx, LUA_TBOOLEAN
|   mov dword tv.value, edx		// Assumes: LUA_TBOOLEAN == 1
|   settt tv, edx
||} else {  /* false */
|   mov dword tv.value, 0
|   settt tv, LUA_TBOOLEAN
||}
|.endmacro
|
|.macro loadnvaluek, vptr
||if ((vptr)->n == (lua_Number)0) {
|  fldz
||} else if ((vptr)->n == (lua_Number)1) {
|  fld1
||} else {
|  fld qword [vptr]
||}
|.endmacro
|
|.macro setnvaluek, tv, vptr		// Pass a Value *! With permanent addr.
|  // SSE2 does not pay off here (I tried).
|  loadnvaluek vptr
|  fstp qword tv.value
|  settt tv, LUA_TNUMBER
|.endmacro
|
|.macro setnvalue, tv, vptr		// Pass a Value *! Temporary ok.
|  mov dword tv.value, (vptr)->na[0]
|  mov dword tv.value.na[1], (vptr)->na[1]
|  settt tv, LUA_TNUMBER
|.endmacro
|
|.macro setsvalue, tv, vptr
|  mov aword tv.value, vptr
|  settt tv, LUA_TSTRING
|.endmacro
|
|.macro sethvalue, tv, vptr
|  mov aword tv.value, vptr
|  settt tv, LUA_TTABLE
|.endmacro
|
|.macro copyslotSSE, D, S, R1		// May use xmm0.
|  mov R1, S.tt;  movq xmm0, qword S.value
|  mov D.tt, R1;  movq qword D.value, xmm0
|.endmacro
|
|.macro copyslot, D, S, R1, R2, R3
||if (J->flags & JIT_F_CPU_SSE2) {
|  copyslotSSE D, S, R1
||} else {
|  mov R1, S.value;  mov R2, S.value.na[1];  mov R3, S.tt
|  mov D.value, R1;  mov D.value.na[1], R2;  mov D.tt, R3
||}
|.endmacro
|
|.macro copyslot, D, S, R1, R2
||if (J->flags & JIT_F_CPU_SSE2) {
|  copyslotSSE D, S, R1
||} else {
|  mov R1, S.value;  mov R2, S.value.na[1];  mov D.value, R1
|  mov R1, S.tt;  mov D.value.na[1], R2;  mov D.tt, R1
||}
|.endmacro
|
|.macro copyslot, D, S
|  copyslot D, S, ecx, edx, eax
|.endmacro
|
|.macro copyconst, tv, tvk		// May use edx.
||switch (ttype(tvk)) {
||case LUA_TNIL:
|   setnilvalue tv
||  break;
||case LUA_TBOOLEAN:
|   setbvalue tv, bvalue(tvk)		// May use edx.
||  break;
||case LUA_TNUMBER: {
|   setnvaluek tv, &(tvk)->value
||  break;
||}
||case LUA_TSTRING:
|   setsvalue tv, &gcvalue(tvk)
||  break;
||default: lua_assert(0); break;
||}
|.endmacro
|
|// Macros to get Lua structures.
|.macro getLCL, reg			// May use CI and TOP (edi).
||if (!J->pt->is_vararg) {
|  mov LCL:reg, BASE[-1].value
||} else {
|  mov CI, L->ci
|  mov TOP, CI->func
|  mov LCL:reg, TOP->value
||}
|.endmacro
|.macro getLCL;  getLCL eax; .endmacro
|
|// Macros to handle variants.
|.macro addidx, type, idx
||if (idx) {
|  add type, idx*#type
||}
|.endmacro
|
|.macro subidx, type, idx
||if (idx) {
|  sub type, idx*#type
||}
|.endmacro
|
|// Annoying x87 stuff: support for two compare variants.
|.macro fcomparepp			// Compare and pop st0 >< st1.
||if (J->flags & JIT_F_CPU_CMOV) {
|  fucomip st1
|  fpop
||} else {
|  fucompp
|  fnstsw ax				// eax modified!
|  sahf
|  // Sometimes test ah, imm8 would be faster.
|  // But all following compares need to be changed then.
|  // Don't bother since this is only compatibility stuff for old CPUs.
||}
|.endmacro
|
|// If you change LUA_TVALUE_ALIGN, be sure to change the Makefile, too:
|//   DASMFLAGS= -D TVALUE_SIZE=...
|// Then rerun make. Or change the default below:
|.if not TVALUE_SIZE; .define TVALUE_SIZE, 16; .endif
|
|.if TVALUE_SIZE == 16
|  .macro TValuemul, reg;  sal reg, 4; .endmacro  // *16
|  .macro TValuediv, reg;  sar reg, 4; .endmacro  // /16
|  .macro Nodemul, reg;  sal reg, 5; .endmacro    // *32
|.elif TVALUE_SIZE == 12
|  .macro TValuemul, reg;  sal reg, 2;  lea reg, [reg+reg*2]; .endmacro  // *12
|  .macro TValuediv, reg;  sal reg, 2;  imul reg, 0xaaaaaaab; .endmacro  // /12
|  .macro Nodemul, reg;  imul reg, 28; .endmacro  // *28
|.else
|  .fatal Unsupported TValue size `TVALUE_SIZE'.
|.endif
|
|//
|// x86 C calling conventions and stack frame layout during a JIT call:
|//
|// ebp+aword*4  CARG2     nresults
|// ebp+aword*3  CARG2     func      (also used as SAVER3 for L)
|// ebp+aword*2  CARG1     L
|// -------------------------------  call to GATE_LJ
|// ebp+aword*1  retaddr
|// ebp+aword*0  frameptr       ebp
|// ebp-aword*1  SAVER1    TOP  edi
|// ebp-aword*2  SAVER2    BASE ebx
|// -------------------------------
|//              GATE_LJ retaddr
|// esp+aword*2  ARG3
|// esp+aword*1  ARG2
|// esp+aword*0  ARG1                <-- esp for first JIT frame
|// -------------------------------
|//              1st JIT frame retaddr
|// esp+aword*2  ARG3
|// esp+aword*1  ARG2
|// esp+aword*0  ARG1                <-- esp for second JIT frame
|// -------------------------------
|//              2nd JIT frame retaddr
|//
|// We could omit the fixed frame pointer (ebp) and have one more register
|// available. But there is no pressing need (could use it for CI).
|// And it's easier for debugging (gdb is still confused -- why?).
|//
|// The stack is aligned to 4 awords (16 bytes). Calls to C functions
|// with up to 3 arguments do not need any stack pointer adjustment.
|//
|
|.define CARG3, [ebp+aword*4]
|.define CARG2, [ebp+aword*3]
|.define CARG1, [ebp+aword*2]
|.define SAVER1, [ebp-aword*1]
|.define SAVER2, [ebp-aword*2]
|.define ARG7, aword [esp+aword*6]	// Requires large frame.
|.define ARG6, aword [esp+aword*5]	// Requires large frame.
|.define ARG5, aword [esp+aword*4]	// Requires large frame.
|.define ARG4, aword [esp+aword*3]	// Requires large frame.
|.define ARG3, aword [esp+aword*2]
|.define ARG2, aword [esp+aword*1]
|.define ARG1, aword [esp]
|.define FRAME_RETADDR, aword [esp+aword*3]
|.define TMP3, [esp+aword*2]
|.define TMP2, [esp+aword*1]
|.define TMP1, [esp]
|.define FPARG2, qword [esp+qword*1]	// Requires large frame.
|.define FPARG1, qword [esp]
|.define LJFRAME_OFFSET, aword*2	// 16 byte aligned with retaddr + ebp.
|.define FRAME_OFFSET, aword*3		// 16 byte aligned with retaddr.
|.define LFRAME_OFFSET, aword*7		// 16 byte aligned with retaddr.
|.define S2LFRAME_OFFSET, aword*4	// Delta to large frame.
|
|.macro call, target, a1
|  mov ARG1, a1;  call target; .endmacro
|.macro call, target, a1, a2
|  mov ARG1, a1;  mov ARG2, a2;  call target; .endmacro
|.macro call, target, a1, a2, a3
|  mov ARG1, a1;  mov ARG2, a2;  mov ARG3, a3;  call target; .endmacro
|.macro call, target, a1, a2, a3, a4
|  push a4;  push a3;  push a2;  push a1
|  call target;  add esp, S2LFRAME_OFFSET;  .endmacro
|.macro call, target, a1, a2, a3, a4, a5
|  mov ARG1, a5; push a4; push a3;  push a2;  push a1
|  call target;  add esp, S2LFRAME_OFFSET;  .endmacro
|
|// The following macros require a large frame.
|.macro call_LFRAME, target, a1, a2, a3, a4
|  mov ARG1, a1;  mov ARG2, a2;  mov ARG3, a3;  mov ARG4, a4
|  call target; .endmacro
|.macro call_LFRAME, target, a1, a2, a3, a4, a5
|  mov ARG1, a1;  mov ARG2, a2;  mov ARG3, a3;  mov ARG4, a4;  mov ARG5, a5
|  call target; .endmacro
|