/*
* Load the following and call epilogue:
* r[ RA_NR_ARGS ] = number of results
* r[ RA_BASE ] = start address of results
*/
void do_ret( struct machine_ops* mop, struct emitter* e, struct frame* f, int vregbase, int nret ){
assert( RA_COUNT <= f->m->nr_reg ); // regargs are passed by register NOT stack
vreg_operand basestack = vreg_to_operand( f, vregbase, true );
/*
* if nret == 0 then prev instruction was call and calls save vregs on stack. Therefore
* only when nret > 0 do we need to save live regs onto stack. Do it first so it can
* use as many temps as it wants before we reserve them for return procedure.
*/
if( nret > 0 )
#if 1
save_frame_limit( mop, e, f, vregbase, nret - 1 );
#else
// TODO: the above is from vregbase - so need to think about this one
jfunc_call( mop, e, f->m, JF_STORE_LOCALS, jf_storelocal_offset( f->m, nret - 1 ), JFUNC_UNLIMITED_STACK, 0 );
#endif
operand rargs[ RA_COUNT ];
prefer_nontemp_acquire_reg( mop, e, f->m, RA_COUNT, rargs );
mop->add( e, f->m, rargs[ RA_BASE ], OP_TARGETREG( basestack.value.base ), OP_TARGETIMMED( basestack.value.offset ) );
if( nret > 0 ) {
mop->move( e, f->m, rargs[ RA_NR_ARGS ], OP_TARGETIMMED( nret - 1 ) );
} else {
mop->sub( e, f->m, rargs[ RA_NR_ARGS ], rargs[ RA_BASE ], OP_TARGETREG( f->m->sp ) );
mop->udiv( e, f->m, rargs[ RA_NR_ARGS ], rargs[ RA_NR_ARGS ], OP_TARGETIMMED( 8 ) );
}
mop->b( e, f->m, LBL_ABS( OP_TARGETIMMED( (uintptr_t)jfunc_addr( e, JF_EPILOGUE ) ) ) );
}
void syn_memsetw( struct machine_ops* mop, struct emitter* e, struct machine* m, operand d, operand v, operand size ){
operand iter = OP_TARGETREG( acquire_temp( mop, e, m ) );
operand dst = OP_TARGETREG( acquire_temp( mop, e, m ) );
// init iterator
mop->move( e, m, iter, OP_TARGETIMMED( 0 ) );
mop->move( e, m, dst, d );
// start loop
e->ops->label_local( e, 0 );
mop->beq( e, m, iter, size, LBL_NEXT( 0 ) );
// copy
mop->move( e, m, OP_TARGETDADDR( dst.reg, 0 ), v );
// update pointers
mop->add( e, m, dst, dst, OP_TARGETIMMED( 4 ) );
// update iterator
mop->add( e, m, iter, iter, OP_TARGETIMMED( 1 ) );
mop->b( e, m, LBL_PREV( 0 ) );
e->ops->label_local( e, 0 );
release_tempn( mop, e, m, 2 );
}
// this function could be inline
void syn_memcpyw( struct machine_ops* mop, struct emitter* e, struct machine* m, operand d, operand s, operand size ){
operand iter = OP_TARGETREG( acquire_temp( mop, e, m ) );
operand src = OP_TARGETREG( acquire_temp( mop, e, m ) );
operand dst = OP_TARGETREG( acquire_temp( mop, e, m ) );
// init iterator
mop->move( e, m, iter, OP_TARGETIMMED( 0 ) );
mop->move( e, m, src, s );
mop->move( e, m, dst, d );
// start loop
e->ops->label_local( e, 0 );
mop->beq( e, m, iter, size, LBL_NEXT( 0 ) );
// copy
mop->move( e, m, OP_TARGETDADDR( dst.reg, 0 ), OP_TARGETDADDR( src.reg, 0 ) );
// update pointers
mop->add( e, m, dst, dst, OP_TARGETIMMED( -4 ) ); // TODO: this is incorrect in general but correct for copyargs
mop->add( e, m, src, src, OP_TARGETIMMED( -4 ) );
// update iterator
mop->add( e, m, iter, iter, OP_TARGETIMMED( 1 ) );
mop->b( e, m, LBL_PREV( 0 ) );
e->ops->label_local( e, 0 );
release_tempn( mop, e, m, 3 );
}
void epilogue( struct machine_ops* mop, struct emitter* e, struct frame* f ){
#if 0 // epilogue is now JFunction - so shared amongst all functions
const operand sp = OP_TARGETREG( f->m->sp );
const operand fp = OP_TARGETREG( f->m->fp );
// reset stack
mop->move( e, f->m, sp, fp );
pop( mop, e, f->m, fp );
mop->ret( e, f->m );
#endif
}
void jinit_cpy_arg_res( struct JFunc* jf, struct machine_ops* mop, struct emitter* e, struct machine* m ){
operand rargs[ RA_SIZE ];
prefer_nontemp_acquire_reg( mop, e, m, RA_SIZE, rargs );
syn_min( mop, e, m, rargs[ RA_EXIST ], rargs[ RA_EXIST ], rargs[ RA_EXPECT ] );
// init iterator
operand iter = OP_TARGETREG( acquire_temp( mop, e, m ) );
mop->move( e, m, iter, OP_TARGETIMMED( 0 ) );
do_copying( mop, e, m, iter, rargs[ RA_EXIST ], rargs[ RA_DST ], rargs[ RA_SRC ] );
/*
* TODO: Below logic is incorrect the first is most likely to get spilled. So change enum order. BUT
* then you have to think about prefer saved reg and the order there.
* TODO: add error prefer_nontemps to error when not all live simultenously
*
* RA_EXPECT is last register therefore its the most likely to be spilled. So to stop repeat
* spill/unspill move it to exist.
*/
mop->move( e, m, rargs[ RA_SIZE ], rargs[ RA_EXIST ] );
do_nilling( mop, e, m, iter, rargs[ RA_EXIST ], rargs[ RA_DST ] );
release_temp( mop, e, m );
prefer_nontemp_release_reg( mop, e, m, RA_SIZE );
mop->ret( e, m );
}
void address_of( struct machine_ops *mop, struct emitter *e, struct machine *m,
operand d,
operand s ){
assert( ISO_DADDR( s ) );
mop->add( e, m, d, OP_TARGETREG( s.base ),
OP_TARGETIMMED( s.offset ) );
}
/* Grab non temps first then temps */
static void prefer_nontemp_acquire_reg( struct machine_ops* mop, struct emitter* e, struct machine* m,
int n, operand reg[n] ){
// must fit into registers only
assert( n <= m->nr_reg );
const int nr_temps = m->nr_temp_regs;
const int nr_nontemps = m->nr_reg - nr_temps;
for( int i = 0; i < n; i++ ){
if( i < nr_nontemps )
reg[i] = OP_TARGETREG( m->reg[ nr_temps + i ] );
else
reg[i] = OP_TARGETREG( acquire_temp( mop, e, m ) ); // ( i - nr_nontemps );
}
// disable spill
disable_spill( m );
}
static void postcall( struct machine_ops* mop, struct emitter* e, struct frame* f, int vregbase, int narg, int nret ){
vreg_operand basestack = vreg_to_operand( f, vregbase, true );
operand rargs[ RA_SIZE ];
prefer_nontemp_acquire_reg( mop, e, f->m, RA_SIZE, rargs );
// max stack clobber
const int maxstack = 3; // prior frame has buffer of pushed return addr, frame pointer and closure
#define USE_JFUNC_FOR_VARRES
// set dst and expect
#if 0
mop->add( e, f->m, rargs[ RA_DST ], OP_TARGETREG( basestack.value.base ), OP_TARGETIMMED( basestack.value.offset ) );
#endif
if( nret == 0 ){
#ifdef USE_JFUNC_FOR_VARRES
jfunc_call( mop, e, f->m, JF_VARRES_POSTCALL, 0, maxstack, 4, rargs[ RA_SRC ], rargs[ RA_DST ],
rargs[ RA_EXPECT ], rargs[ RA_EXIST ] );
#else
mop->move( e, f->m, rargs[ RA_EXPECT ], rargs[ RA_EXIST ] );
// update stack position
mop->mul( e, f->m, rargs[ RA_NR_ARGS ], rargs[ RA_NR_ARGS ], OP_TARGETIMMED( 8 ) ); // in word units
mop->add( e, f->m, rargs[ RA_NR_ARGS ], rargs[ RA_NR_ARGS ], OP_TARGETIMMED( 8 + 8 * vregbase ) );
mop->sub( e, f->m, OP_TARGETREG( f->m->sp ), OP_TARGETREG( f->m->fp ), rargs[ RA_NR_ARGS ] );
mop->move( e, f->m, rargs[ RA_EXIST ], rargs[ RA_EXPECT ] );
#endif
} else {
mop->move( e, f->m, rargs[ RA_EXPECT ], OP_TARGETIMMED( nret - 1 ) );
#ifdef USE_JFUNC_FOR_VARRES
jfunc_call( mop, e, f->m, JF_ARG_RES_CPY, 0, maxstack, 4, rargs[ RA_SRC ], rargs[ RA_DST ],
rargs[ RA_EXPECT ], rargs[ RA_EXIST ] );
#endif
}
#ifndef USE_JFUNC_FOR_VARRES
jfunc_call( mop, e, f->m, JF_ARG_RES_CPY, 0, maxstack, 4, rargs[ RA_SRC ], rargs[ RA_DST ],
rargs[ RA_EXPECT ], rargs[ RA_EXIST ] );
#endif
prefer_nontemp_release_reg( mop, e, f->m, RA_SIZE );
}
static void precall( struct machine_ops* mop, struct emitter* e, struct frame* f, int vregbase, int narg, int nret ){
// new frame assumes no temporaries have been used yet
assert( temps_accessed( f->m ) == 0 );
assert( RA_COUNT + 1 <= f->m->nr_reg ); // regargs are passed by register NOT stack
vreg_operand clive = vreg_to_operand( f, vregbase, false );
vreg_operand cstack = vreg_to_operand( f, vregbase, true );
assert( cstack.value.tag == OT_DIRECTADDR );
// TODO: verify its a closure using clive
// get arg passing registers
operand rargs[ RA_COUNT ];
prefer_nontemp_acquire_reg( mop, e, f->m, RA_COUNT, rargs );
// calculate number of args
if( narg > 0 )
mop->move( e, f->m, rargs[ RA_NR_ARGS ], OP_TARGETIMMED( narg - 1 ) );
else{
// calculate total by subtracting basereg address from stack.
// 2 becuase 8 for (ebp,closure) another 8 for the function being called ( rem actual args = args - 1 )
mop->add( e, f->m, rargs[ RA_NR_ARGS ], OP_TARGETREG( f->m->fp ), OP_TARGETIMMED( -8 * ( 2 + vregbase ) ) );
mop->sub( e, f->m, rargs[ RA_NR_ARGS ], rargs[ RA_NR_ARGS ], OP_TARGETREG( f->m->sp ) );
mop->udiv( e, f->m, rargs[ RA_NR_ARGS ], rargs[ RA_NR_ARGS ], OP_TARGETIMMED( 8 ) );
}
// calcualte base address
mop->add( e, f->m, rargs[ RA_BASE ], OP_TARGETREG( cstack.value.base ), OP_TARGETIMMED( cstack.value.offset ) );
// call function without spilling any temps
#if 0
bool prior = disable_spill( f->m );
mop->call( e, f->m, LBL_ABS( clive.value ) );
restore_spill( f->m, prior );
#else
jfunc_call( mop, e, f->m, JF_PROLOGUE, 0, JFUNC_UNLIMITED_STACK, 2, rargs[ RA_EXIST ], rargs[ RA_SRC ] );
#endif
// release temps used in call
prefer_nontemp_release_reg( mop, e, f->m, RA_COUNT );
}
static operand pslot_to_operand( struct frame* f, int nr_locals, int pidx, bool stackonly ){
assert( f );
assert( f->m );
if( stackonly || pidx >= NR_SPARE_REGS( f ) ){
#if 0
operand r = OP_TARGETDADDR( f->m->sp, 4 * INVERSE( pidx, 2 * nr_locals ) );
#else
operand r = OP_TARGETDADDR( f->m->fp, -(8 + 4 * pidx) );
#endif
return r;
} else {
operand r = OP_TARGETREG( vreg_to_physical_reg( f, pidx ) );
return r;
}
}
void popn( struct machine_ops* mop, struct emitter* e, struct machine* m, int nr_operands, ... ){
va_list ap;
const operand stack = OP_TARGETREG( m->sp );
va_start( ap, nr_operands );
if( nr_operands == 1 && mop->pop ){
mop->pop( e, m, va_arg( ap, operand ) );
} else {
for( int i = 0; i < nr_operands; i++ )
mop->move( e, m, va_arg( ap, operand ), OP_TARGETDADDR( m->sp, 4 * i ) );
mop->add( e, m, stack, stack, OP_TARGETIMMED( 4 * nr_operands ) );
}
va_end( ap );
}
void mips_b( struct emitter* me, struct machine* m, label l ){
if( !ISL_ABS( l ) ){
EMIT( MI_B( mips_branch( me, l ) ) );
} else if ( ISL_ABSDIRECT( l ) ){ // TODO: make sure within 256MB region by calling MI_J_SAFE
EMIT( MI_J( l.abs.k ) );
} else if ( ISO_REG( l.abs ) ) {
EMIT( MI_JR( l.abs.reg ) );
} else {
assert( ISO_DADDR( l.abs ) );
operand t = OP_TARGETREG( acquire_temp( _MOP, me, m ) );
move( me, m, t, l.abs );
EMIT( MI_JR( t.reg ) );
RELEASE_OR_NOP( me, m );
return;
}
// delay slot
EMIT( MI_NOP() );
}
static lua_Number ljc_relational( lua_Number st, lua_Number sv
, lua_Number tt, lua_Number tv
, int op ) {
assert( !( st == LUA_TNUMBER && tt == LUA_TNUMBER ) );
struct TValue s = { .t = st, .v = (union Value)sv };
struct TValue t = { .t = tt, .v = (union Value)tv };
switch( op ){
case REL_LT:
return do_lt( &s, &t );
case REL_LEQ:
return do_leq( &s, &t );
case REL_EQ:
return do_eq( &s, &t );
default:
assert( false );
}
}
typedef void (*arch_rel)( struct emitter*, struct machine*
, operand, operand, label );
static void emit_relational( struct emitter *me, struct machine_ops *mop
, struct frame* f
, loperand s, loperand t
, arch_rel ar, int op
, bool expect ){
vreg_operand os = loperand_to_operand( f, s ),
ot = loperand_to_operand( f, t );
unsigned int pc = me->ops->pc( me ) + 2;
label l = LBL_PC( pc );
// determine if coercion is required
operand tag = OP_TARGETREG( acquire_temp( mop, me, f->m ) );
mop->bor( me, f->m, tag, os.type, ot.type );
mop->beq( me, f->m, tag, OP_TARGETIMMED( 0 ), LBL_NEXT( 0 ) );
// do coercion
mop->call_static_cfn( me, f, (uintptr_t)&ljc_relational
, &tag, 5, os.type, os.value
, ot.type, ot.value
, OP_TARGETIMMED( op ) );
mop->beq( me, f->m, tag, OP_TARGETIMMED( expect ), l );
mop->b( me, f->m, LBL_NEXT( 1 ) );
// do primitive relational
me->ops->label_local( me, 0 );
ar( me, f->m, os.value, ot.value, l );
me->ops->label_local( me, 1 );
release_temp( mop, me, f->m );
return;
}
void emit_jmp( struct emitter** mce, struct machine_ops* mop
, struct frame *f
, loperand a
, int offset ){
assert( a.islocal );
// if not zero then any upvalues below the vreg need to be closed.
if( a.index > 0 ){
vreg_operand op = vreg_to_operand( f, a.index + 1, true );
operand base = OP_TARGETREG( acquire_temp( mop, REF, f->m ) );
address_of( mop, REF, f->m, base, op.type );
mop->call_static_cfn( REF, f, (uintptr_t)&closure_close, NULL
, 1
, base );
release_temp( mop, REF, f->m );
}
unsigned int pc = (int)REF->ops->pc( REF ) + offset + 1;
mop->b( REF, f->m, LBL_PC( pc ) );
}
void jinit_epi( struct JFunc* jf, struct machine_ops* mop, struct emitter* e, struct machine* m ){
// phoney frame
struct frame F = { .m = m, .nr_locals = 1, .nr_params = 0 };
struct frame *f = &F;
const operand sp = OP_TARGETREG( m->sp );
const operand fp = OP_TARGETREG( m->fp );
operand rargs[ RA_SIZE ];
prefer_nontemp_acquire_reg( mop, e, f->m, RA_SIZE, rargs );
// reset stack
// mop->move( e, m, sp, fp );
mop->add( e, m, sp, fp, OP_TARGETIMMED( -4 ) );
if( m->is_ra )
popn( mop, e, m , 3, rargs[ RA_DST ], fp, OP_TARGETREG( m->ra ) );
else
popn( mop, e, m, 2, rargs[ RA_DST ], fp );
// pop( mop, e, m, fp );
mop->ret( e, m );
prefer_nontemp_release_reg( mop, e, f->m, RA_SIZE );
}
/*
* Do the majority ( function independent ) part of the prologue. That is: store the frame section,
* update src and dst pointers and call the memcpy.
*
* The function specific code needs to set the number of params, update the stack ( requires # of locals )
* and then unspill params.
*/
void jinit_pro( struct JFunc* jf, struct machine_ops* mop, struct emitter* e, struct machine* m ){
// phoney frame
struct frame F = { .m = m, .nr_locals = 1, .nr_params = 0 };
struct frame *f = &F;
const operand sp = OP_TARGETREG( f->m->sp );
const operand fp = OP_TARGETREG( f->m->fp );
const vreg_operand basestack = vreg_to_operand( f, 0, true ); // destination is first local
const int maxstack = JFUNC_UNLIMITED_STACK;
operand rargs[ RA_SIZE ];
prefer_nontemp_acquire_reg( mop, e, f->m, RA_SIZE, rargs );
// push old frame pointer, closure addr / result start addr, expected nr or results
if( f->m->is_ra )
pushn( mop, e, f->m, 3, OP_TARGETREG( f->m->ra), fp, rargs[ RA_SRC ] );
else
pushn( mop, e, f->m, 2, fp, rargs[ RA_SRC ] );
// set ebp and update stack
mop->add( e, f->m, fp, sp, OP_TARGETIMMED( 4 ) ); // point to ebp so add 4
// set src ( always start after closure see Lua VM for reason )
mop->add( e, f->m, rargs[ RA_SRC ], rargs[ RA_SRC ], OP_TARGETIMMED( -8 ) );
mop->add( e, f->m, rargs[ RA_DST ], OP_TARGETREG( basestack.value.base ), OP_TARGETIMMED( basestack.value.offset ) );
/*
* Call the actual function, which is the closure. On RISC this will clobber
* temp hopefully this isn't a live reg or we will get exception. On CISC
* there is probably indirect direct address jmp instruction ( x86 does 0 ).
*/
mop->b( e, f->m, LBL_ABS( OP_TARGETDADDR( rargs[ RA_SRC ].reg, 8 ) ) );
prefer_nontemp_release_reg( mop, e, f->m, RA_SIZE );
}
/*
* The number of results is not know before hand. Need to update stack for future
* calls.
*/
void jinit_vresult_postcall( struct JFunc* jf, struct machine_ops* mop, struct emitter* e, struct machine* m ){
// phoney frame
struct frame F = { .m = m, .nr_locals = 1, .nr_params = 0 };
struct frame *f = &F;
operand rargs[ RA_SIZE ];
prefer_nontemp_acquire_reg( mop, e, f->m, RA_SIZE, rargs );
// max stack clobber
const int maxstack = 3; // prior frame has buffer of pushed return addr, frame pointer and closure
// consume as many results as available
mop->move( e, f->m, rargs[ RA_EXPECT ], rargs[ RA_EXIST ] );
// if register based remember return address
if( f->m->is_ra )
pushn( mop, e, f->m, 1, OP_TARGETREG( f->m->ra) ); // not safe cause of stack
// copy args across
jfunc_call( mop, e, f->m, JF_ARG_RES_CPY, 0, maxstack, 4, rargs[ RA_SRC ], rargs[ RA_DST ],
rargs[ RA_EXPECT ], rargs[ RA_EXIST ] );
if( f->m->is_ra )
popn( mop, e, f->m, 1, OP_TARGETREG( f->m->ra) );
/*
* this depends heavily on copy arg implementation, it assumes ptrs will point to
* the top of the stack after copying i.e. the last result copied.
*/
mop->add( e, m, OP_TARGETREG( f->m->sp ), rargs[ RA_DST ], OP_TARGETIMMED( 0 ) );
prefer_nontemp_release_reg( mop, e, f->m, RA_SIZE );
// return
mop->ret( e, m );
}
void prologue( struct machine_ops* mop, struct emitter* e, struct frame* f ){
#if 1
const operand sp = OP_TARGETREG( f->m->sp );
const operand fp = OP_TARGETREG( f->m->fp );
const int nparams = f->nr_params;
operand rargs[ RA_SIZE ];
prefer_nontemp_acquire_reg( mop, e, f->m, RA_SIZE, rargs );
mop->add( e, f->m, sp, sp, OP_TARGETIMMED( -( 8 * f->nr_locals ) ) );
if( nparams ){
// set nparams
mop->move( e, f->m, rargs[ RA_EXPECT ], OP_TARGETIMMED( nparams ) );
// do argument cpy
jfunc_call( mop, e, f->m, JF_ARG_RES_CPY, 0, JFUNC_UNLIMITED_STACK, 4, rargs[ RA_SRC ], rargs[ RA_DST ],
rargs[ RA_EXPECT ], rargs[ RA_EXIST ] );
// do call
jfunc_call( mop, e, f->m, JF_LOAD_LOCALS, jf_loadlocal_offset( f->m, nparams ), JFUNC_UNLIMITED_STACK, 0 );
}
prefer_nontemp_release_reg( mop, e, f->m, RA_SIZE );
#else
// new frame assumes no temporaries have been used yet
assert( temps_accessed( f->m ) == 0 );
const operand sp = OP_TARGETREG( f->m->sp );
const operand fp = OP_TARGETREG( f->m->fp );
const int nparams = f->nr_params;
operand rargs[ RA_SIZE ];
prefer_nontemp_acquire_reg( mop, e, f->m, RA_SIZE, rargs );
// push old frame pointer, closure addr / result start addr, expected nr or results
if( f->m->is_ra )
pushn( mop, e, f->m, 3, OP_TARGETREG( f->m->ra), fp, rargs[ RA_SRC ] );
else
pushn( mop, e, f->m, 2, fp, rargs[ RA_SRC ] );
// set ebp and update stack
mop->add( e, f->m, fp, sp, OP_TARGETIMMED( 4 ) ); // point to ebp so add 4
mop->add( e, f->m, sp, sp, OP_TARGETIMMED( -( 8 * f->nr_locals ) ) );
if( nparams ) {
const vreg_operand basestack = vreg_to_operand( f, 0, true ); // destination is first local
const int maxstack = JFUNC_UNLIMITED_STACK;
// set src ( always start after closure see Lua VM for reason )
mop->add( e, f->m, rargs[ RA_SRC ], rargs[ RA_SRC ], OP_TARGETIMMED( -8 ) );
// set dst and expect
mop->add( e, f->m, rargs[ RA_DST ], OP_TARGETREG( basestack.value.base ), OP_TARGETIMMED( basestack.value.offset ) );
mop->move( e, f->m, rargs[ RA_EXPECT ], OP_TARGETIMMED( nparams ) );
jfunc_call( mop, e, f->m, JF_ARG_RES_CPY, 0, maxstack, 4, rargs[ RA_SRC ], rargs[ RA_DST ],
rargs[ RA_EXPECT ], rargs[ RA_EXIST ] );
prefer_nontemp_release_reg( mop, e, f->m, RA_SIZE );
#if 0
load_frame_limit( mop, e, f, 0, nparams ); // load locals living in registers
#else
jfunc_call( mop, e, f->m, JF_LOAD_LOCALS, jf_loadlocal_offset( f->m, nparams ), maxstack, 0 );
#endif
}
#endif
}
void loadim( struct machine_ops* mop, struct emitter* e, struct machine* m, int reg, int value ) {
operand r = OP_TARGETREG( reg );
operand v = OP_TARGETIMMED( value );
mop->move( e, m, r, v );
}