static Node *slicebase(Simp *s, Node *n, Node *off)
{
Node *t, *u, *v;
Type *ty;
int sz;
t = rval(s, n, NULL);
u = NULL;
ty = tybase(exprtype(n));
switch (ty->type) {
case Typtr: u = t; break;
case Tyarray: u = addr(s, t, base(exprtype(n))); break;
case Tyslice: u = load(addr(s, t, mktyptr(n->line, base(exprtype(n))))); break;
default: die("Unslicable type %s", tystr(n->expr.type));
}
/* safe: all types we allow here have a sub[0] that we want to grab */
if (off) {
off = ptrsized(s, rval(s, off, NULL));
sz = tysize(n->expr.type->sub[0]);
v = mul(off, disp(n->line, sz));
return add(u, v);
} else {
return u;
}
}
targ_size_t size(tym_t ty)
{
int sz = (tybasic(ty) == TYvoid) ? 1 : tysize(ty);
#ifdef DEBUG
if (sz == -1)
WRTYxx(ty);
#endif
assert(sz!= -1);
return sz;
}
size_t size(Node *n)
{
Type *t;
if (n->type == Nexpr)
t = n->expr.type;
else
t = n->decl.type;
return tysize(t);
}
bool xmmIsAligned(elem *e)
{
if (tyvector(e->Ety) && e->Eoper == OPvar)
{
Symbol *s = e->EV.sp.Vsym;
if (s->Salignsize() < 16 ||
e->EV.sp.Voffset & (16 - 1) ||
tysize(e->Ety) > STACKALIGN
)
return false; // definitely not aligned
}
return true; // assume aligned
}
static Node *simpucon(Simp *s, Node *n, Node *dst)
{
Node *tmp, *u, *tag, *elt, *sz;
Node *r;
Type *ty;
Ucon *uc;
size_t i;
/* find the ucon we're constructing here */
ty = tybase(n->expr.type);
uc = NULL;
for (i = 0; i < ty->nmemb; i++) {
if (!strcmp(namestr(n->expr.args[0]), namestr(ty->udecls[i]->name))) {
uc = ty->udecls[i];
break;
}
}
if (!uc)
die("Couldn't find union constructor");
if (dst)
tmp = dst;
else
tmp = temp(s, n);
/* Set the tag on the ucon */
u = addr(s, tmp, mktype(n->line, Tyuint));
tag = mkintlit(n->line, uc->id);
tag->expr.type = mktype(n->line, Tyuint);
append(s, set(deref(u), tag));
/* fill the value, if needed */
if (!uc->etype)
return tmp;
elt = rval(s, n->expr.args[1], NULL);
u = addk(u, Wordsz);
if (stacktype(uc->etype)) {
elt = addr(s, elt, uc->etype);
sz = disp(n->line, tysize(uc->etype));
r = mkexpr(n->line, Oblit, u, elt, sz, NULL);
} else {
r = set(deref(u), elt);
}
append(s, r);
return tmp;
}
unsigned xmmload(tym_t tym, bool aligned)
{ unsigned op;
if (tysize(tym) == 32)
aligned = false;
switch (tybasic(tym))
{
case TYuint:
case TYint:
case TYlong:
case TYulong: op = LODD; break; // MOVD
case TYfloat:
case TYcfloat:
case TYifloat: op = LODSS; break; // MOVSS
case TYllong:
case TYullong: op = LODQ; break; // MOVQ
case TYdouble:
case TYcdouble:
case TYidouble: op = LODSD; break; // MOVSD
case TYfloat8:
case TYfloat4: op = aligned ? LODAPS : LODUPS; break; // MOVAPS / MOVUPS
case TYdouble4:
case TYdouble2: op = aligned ? LODAPD : LODUPD; break; // MOVAPD / MOVUPD
case TYschar16:
case TYuchar16:
case TYshort8:
case TYushort8:
case TYlong4:
case TYulong4:
case TYllong2:
case TYullong2:
case TYschar32:
case TYuchar32:
case TYshort16:
case TYushort16:
case TYlong8:
case TYulong8:
case TYllong4:
case TYullong4: op = aligned ? LODDQA : LODDQU; break; // MOVDQA / MOVDQU
default:
printf("tym = x%x\n", tym);
assert(0);
}
return op;
}
static Node *intconvert(Simp *s, Node *from, Type *to, int issigned)
{
Node *r;
size_t fromsz, tosz;
fromsz = size(from);
tosz = tysize(to);
r = rval(s, from, NULL);
if (fromsz > tosz) {
r = mkexpr(from->line, Otrunc, r, NULL);
} else if (tosz > fromsz) {
if (issigned)
r = mkexpr(from->line, Oswiden, r, NULL);
else
r = mkexpr(from->line, Ozwiden, r, NULL);
}
r->expr.type = to;
return r;
}
static Node *idxaddr(Simp *s, Node *seq, Node *idx)
{
Node *a, *t, *u, *v; /* temps */
Node *r; /* result */
Type *ty;
size_t sz;
a = rval(s, seq, NULL);
ty = exprtype(seq)->sub[0];
if (exprtype(seq)->type == Tyarray)
t = addr(s, a, ty);
else if (seq->expr.type->type == Tyslice)
t = load(addr(s, a, mktyptr(seq->line, ty)));
else
die("Can't index type %s\n", tystr(seq->expr.type));
assert(t->expr.type->type == Typtr);
u = rval(s, idx, NULL);
u = ptrsized(s, u);
sz = tysize(ty);
v = mul(u, disp(seq->line, sz));
r = add(t, v);
return r;
}
static void sliceStructs_Replace(SymInfo *sia, elem *e)
{
while (1)
{
switch (e->Eoper)
{
case OPvar:
{
Symbol *s = e->EV.sp.Vsym;
SYMIDX si = s->Ssymnum;
//printf("e: %d %d\n", si, sia[si].canSlice);
//elem_print(e);
if (si >= 0 && sia[si].canSlice)
{
if (tysize(e->Ety) == 2 * REGSIZE)
{
// Rewrite e as (si0 OPpair si0+1)
elem *e1 = el_calloc();
el_copy(e1, e);
e1->Ety = sia[si].ty0;
elem *e2 = el_calloc();
el_copy(e2, e);
Symbol *s1 = globsym.tab[sia[si].si0 + 1]; // +1 for second slice
e2->Ety = sia[si].ty1;
e2->EV.sp.Vsym = s1;
e2->Eoffset = 0;
e->Eoper = OPpair;
e->E1 = e1;
e->E2 = e2;
}
else if (e->Eoffset == 0) // the first slice of the symbol is the same as the original
{
}
else
{
Symbol *s1 = globsym.tab[sia[si].si0 + 1]; // +1 for second slice
e->EV.sp.Vsym = s1;
e->Eoffset = 0;
//printf("replaced with:\n");
//elem_print(e);
}
}
return;
}
default:
if (OTunary(e->Eoper))
{
e = e->E1;
break;
}
if (OTbinary(e->Eoper))
{
sliceStructs_Replace(sia, e->E2);
e = e->E1;
break;
}
return;
}
}
}
size_t tysize(Type *t)
{
size_t sz;
size_t i;
sz = 0;
if (!t)
die("size of empty type => bailing.");
switch (t->type) {
case Tyvoid:
die("void has no size");
return 1;
case Tybool: case Tyint8:
case Tybyte: case Tyuint8:
return 1;
case Tyint16: case Tyuint16:
return 2;
case Tyint: case Tyint32:
case Tyuint: case Tyuint32:
case Tychar: /* utf32 */
return 4;
case Typtr: case Tyfunc:
case Tyvalist: /* ptr to first element of valist */
return Ptrsz;
case Tyint64: case Tylong:
case Tyuint64: case Tyulong:
return 8;
/*end integer types*/
case Tyfloat32:
return 4;
case Tyfloat64:
return 8;
case Tyslice:
return 2*Ptrsz; /* len; ptr */
case Tyname:
return tysize(t->sub[0]);
case Tyarray:
t->asize = fold(t->asize, 1);
assert(exprop(t->asize) == Olit);
return t->asize->expr.args[0]->lit.intval * tysize(t->sub[0]);
case Tytuple:
for (i = 0; i < t->nsub; i++) {
sz = tyalign(sz, tysize(t->sub[i]));
sz += tysize(t->sub[i]);
}
for (i = 0; i < t->nsub; i++)
sz = tyalign(sz, tysize(t->sub[i]));
return sz;
break;
case Tystruct:
for (i = 0; i < t->nmemb; i++) {
sz = tyalign(sz, size(t->sdecls[i]));
sz += size(t->sdecls[i]);
}
/* the whole struct size should match the biggest alignment */
for (i = 0; i < t->nmemb; i++)
sz = tyalign(sz, size(t->sdecls[i]));
return sz;
break;
case Tyunion:
sz = Wordsz;
for (i = 0; i < t->nmemb; i++)
if (t->udecls[i]->etype)
sz = max(sz, tysize(t->udecls[i]->etype) + Wordsz);
return align(sz, Ptrsz);
break;
case Tybad: case Tyvar: case Typaram: case Tyunres: case Ntypes:
die("Type %s does not have size; why did it get down to here?", tystr(t));
break;
}
return -1;
}
targ_size_t type_size(type *t)
{ targ_size_t s;
unsigned long u;
tym_t tyb;
type_debug(t);
tyb = tybasic(t->Tty);
#ifdef DEBUG
if (tyb >= TYMAX)
/*type_print(t),*/
dbg_printf("tyb = x%lx\n",tyb);
#endif
assert(tyb < TYMAX);
s = tysize[tyb];
if (s == (targ_size_t) -1)
{
switch (tyb)
{
// in case program plays games with function pointers
case TYffunc:
case TYfpfunc:
#if TX86
case TYnfunc: /* in case program plays games with function pointers */
case TYhfunc:
case TYnpfunc:
case TYnsfunc:
case TYfsfunc:
case TYf16func:
case TYifunc:
case TYjfunc:
#endif
#if SCPP
if (ANSI)
synerr(EM_unknown_size,"function"); /* size of function is not known */
#endif
s = 1;
break;
case TYarray:
if (t->Tflags & TFsizeunknown)
{
#if SCPP
synerr(EM_unknown_size,"array"); /* size of array is unknown */
#endif
t->Tflags &= ~TFsizeunknown;
}
if (t->Tflags & TFvla)
{
s = tysize[pointertype];
break;
}
s = type_size(t->Tnext);
u = t->Tdim * (unsigned long) s;
#if TX86 && SCPP
type_chksize(u);
#endif
s = u;
break;
case TYstruct:
t = t->Ttag->Stype; /* find main instance */
/* (for const struct X) */
if (t->Tflags & TFsizeunknown)
{
#if SCPP
template_instantiate_forward(t->Ttag);
if (t->Tflags & TFsizeunknown)
synerr(EM_unknown_size,t->Tty & TYstruct ? prettyident(t->Ttag) : "struct");
t->Tflags &= ~TFsizeunknown;
#endif
}
assert(t->Ttag);
s = t->Ttag->Sstruct->Sstructsize;
break;
#if SCPP
case TYenum:
if (t->Ttag->Senum->SEflags & SENforward)
synerr(EM_unknown_size, prettyident(t->Ttag));
s = type_size(t->Tnext);
break;
#endif
case TYvoid:
#if SCPP && TARGET_WINDOS // GNUC allows it, so we will, too
synerr(EM_void_novalue); // voids have no value
#endif
s = 1;
break;
#if SCPP
case TYref:
case TYmemptr:
case TYvtshape:
s = tysize(tym_conv(t));
break;
case TYident:
synerr(EM_unknown_size, t->Tident);
s = 1;
break;
#endif
#if MARS
case TYref:
s = tysize(TYnptr);
break;
#endif
default:
#ifdef DEBUG
WRTYxx(t->Tty);
#endif
assert(0);
}
}
return s;
}
void FuncDeclaration::toObjFile(int multiobj)
{
FuncDeclaration *func = this;
ClassDeclaration *cd = func->parent->isClassDeclaration();
int reverse;
int has_arguments;
//printf("FuncDeclaration::toObjFile(%p, %s.%s)\n", func, parent->toChars(), func->toChars());
//if (type) printf("type = %s\n", func->type->toChars());
#if 0
//printf("line = %d\n",func->getWhere() / LINEINC);
EEcontext *ee = env->getEEcontext();
if (ee->EEcompile == 2)
{
if (ee->EElinnum < (func->getWhere() / LINEINC) ||
ee->EElinnum > (func->endwhere / LINEINC)
)
return; // don't compile this function
ee->EEfunc = func->toSymbol();
}
#endif
if (semanticRun >= PASSobj) // if toObjFile() already run
return;
// If errors occurred compiling it, such as bugzilla 6118
if (type && type->ty == Tfunction && ((TypeFunction *)type)->next->ty == Terror)
return;
if (!func->fbody)
{
return;
}
if (func->isUnitTestDeclaration() && !global.params.useUnitTests)
return;
if (multiobj && !isStaticDtorDeclaration() && !isStaticCtorDeclaration())
{ obj_append(this);
return;
}
assert(semanticRun == PASSsemantic3done);
semanticRun = PASSobj;
if (global.params.verbose)
printf("function %s\n",func->toChars());
Symbol *s = func->toSymbol();
func_t *f = s->Sfunc;
#if TARGET_WINDOS
/* This is done so that the 'this' pointer on the stack is the same
* distance away from the function parameters, so that an overriding
* function can call the nested fdensure or fdrequire of its overridden function
* and the stack offsets are the same.
*/
if (isVirtual() && (fensure || frequire))
f->Fflags3 |= Ffakeeh;
#endif
#if TARGET_OSX
s->Sclass = SCcomdat;
#else
s->Sclass = SCglobal;
#endif
for (Dsymbol *p = parent; p; p = p->parent)
{
if (p->isTemplateInstance())
{
s->Sclass = SCcomdat;
break;
}
}
/* Vector operations should be comdat's
*/
if (isArrayOp)
s->Sclass = SCcomdat;
if (isNested())
{
// if (!(config.flags3 & CFG3pic))
// s->Sclass = SCstatic;
f->Fflags3 |= Fnested;
}
else
{
const char *libname = (global.params.symdebug)
? global.params.debuglibname
: global.params.defaultlibname;
// Pull in RTL startup code
if (func->isMain())
{ objextdef("_main");
#if TARGET_LINUX || TARGET_OSX || TARGET_FREEBSD || TARGET_OPENBSD || TARGET_SOLARIS
obj_ehsections(); // initialize exception handling sections
#endif
#if TARGET_WINDOS
objextdef("__acrtused_con");
#endif
obj_includelib(libname);
s->Sclass = SCglobal;
}
else if (strcmp(s->Sident, "main") == 0 && linkage == LINKc)
{
#if TARGET_WINDOS
objextdef("__acrtused_con"); // bring in C startup code
obj_includelib("snn.lib"); // bring in C runtime library
#endif
s->Sclass = SCglobal;
}
else if (func->isWinMain())
{
objextdef("__acrtused");
obj_includelib(libname);
s->Sclass = SCglobal;
}
// Pull in RTL startup code
else if (func->isDllMain())
{
objextdef("__acrtused_dll");
obj_includelib(libname);
s->Sclass = SCglobal;
}
}
cstate.CSpsymtab = &f->Flocsym;
// Find module m for this function
Module *m = NULL;
for (Dsymbol *p = parent; p; p = p->parent)
{
m = p->isModule();
if (m)
break;
}
IRState irs(m, func);
Dsymbols deferToObj; // write these to OBJ file later
irs.deferToObj = &deferToObj;
TypeFunction *tf;
enum RET retmethod;
symbol *shidden = NULL;
Symbol *sthis = NULL;
tym_t tyf;
tyf = tybasic(s->Stype->Tty);
//printf("linkage = %d, tyf = x%x\n", linkage, tyf);
reverse = tyrevfunc(s->Stype->Tty);
assert(func->type->ty == Tfunction);
tf = (TypeFunction *)(func->type);
has_arguments = (tf->linkage == LINKd) && (tf->varargs == 1);
retmethod = tf->retStyle();
if (retmethod == RETstack)
{
// If function returns a struct, put a pointer to that
// as the first argument
::type *thidden = tf->next->pointerTo()->toCtype();
char hiddenparam[5+4+1];
static int hiddenparami; // how many we've generated so far
sprintf(hiddenparam,"__HID%d",++hiddenparami);
shidden = symbol_name(hiddenparam,SCparameter,thidden);
shidden->Sflags |= SFLtrue | SFLfree;
#if DMDV1
if (func->nrvo_can && func->nrvo_var && func->nrvo_var->nestedref)
#else
if (func->nrvo_can && func->nrvo_var && func->nrvo_var->nestedrefs.dim)
#endif
type_setcv(&shidden->Stype, shidden->Stype->Tty | mTYvolatile);
irs.shidden = shidden;
this->shidden = shidden;
}
if (vthis)
{
assert(!vthis->csym);
sthis = vthis->toSymbol();
irs.sthis = sthis;
if (!(f->Fflags3 & Fnested))
f->Fflags3 |= Fmember;
}
Symbol **params;
unsigned pi;
// Estimate number of parameters, pi
pi = (v_arguments != NULL);
if (parameters)
pi += parameters->dim;
// Allow extra 2 for sthis and shidden
params = (Symbol **)alloca((pi + 2) * sizeof(Symbol *));
// Get the actual number of parameters, pi, and fill in the params[]
pi = 0;
if (v_arguments)
{
params[pi] = v_arguments->toSymbol();
pi += 1;
}
if (parameters)
{
for (size_t i = 0; i < parameters->dim; i++)
{ VarDeclaration *v = parameters->tdata()[i];
if (v->csym)
{
error("compiler error, parameter '%s', bugzilla 2962?", v->toChars());
assert(0);
}
params[pi + i] = v->toSymbol();
}
pi += parameters->dim;
}
if (reverse)
{ // Reverse params[] entries
for (size_t i = 0; i < pi/2; i++)
{
Symbol *sptmp = params[i];
params[i] = params[pi - 1 - i];
params[pi - 1 - i] = sptmp;
}
}
if (shidden)
{
#if 0
// shidden becomes last parameter
params[pi] = shidden;
#else
// shidden becomes first parameter
memmove(params + 1, params, pi * sizeof(params[0]));
params[0] = shidden;
#endif
pi++;
}
if (sthis)
{
#if 0
// sthis becomes last parameter
params[pi] = sthis;
#else
// sthis becomes first parameter
memmove(params + 1, params, pi * sizeof(params[0]));
params[0] = sthis;
#endif
pi++;
}
if ((global.params.isLinux || global.params.isOSX || global.params.isFreeBSD || global.params.isSolaris) &&
linkage != LINKd && shidden && sthis)
{
/* swap shidden and sthis
*/
Symbol *sp = params[0];
params[0] = params[1];
params[1] = sp;
}
for (size_t i = 0; i < pi; i++)
{ Symbol *sp = params[i];
sp->Sclass = SCparameter;
sp->Sflags &= ~SFLspill;
sp->Sfl = FLpara;
symbol_add(sp);
}
// Determine register assignments
if (pi)
{
if (global.params.is64bit)
{
// Order of assignment of pointer or integer parameters
static const unsigned char argregs[6] = { DI,SI,DX,CX,R8,R9 };
int r = 0;
int xmmcnt = XMM0;
for (size_t i = 0; i < pi; i++)
{ Symbol *sp = params[i];
tym_t ty = tybasic(sp->Stype->Tty);
// BUG: doesn't work for structs
if (r < sizeof(argregs)/sizeof(argregs[0]))
{
if (type_jparam(sp->Stype))
{
sp->Sclass = SCfastpar;
sp->Spreg = argregs[r];
sp->Sfl = FLauto;
++r;
}
}
if (xmmcnt <= XMM7)
{
if (tyfloating(ty) && tysize(ty) <= 8)
{
sp->Sclass = SCfastpar;
sp->Spreg = xmmcnt;
sp->Sfl = FLauto;
++xmmcnt;
}
}
}
}
else
{
// First parameter goes in register
Symbol *sp = params[0];
if ((tyf == TYjfunc || tyf == TYmfunc) &&
type_jparam(sp->Stype))
{ sp->Sclass = SCfastpar;
sp->Spreg = (tyf == TYjfunc) ? AX : CX;
sp->Sfl = FLauto;
//printf("'%s' is SCfastpar\n",sp->Sident);
}
}
}
if (func->fbody)
{ block *b;
Blockx bx;
Statement *sbody;
localgot = NULL;
sbody = func->fbody;
memset(&bx,0,sizeof(bx));
bx.startblock = block_calloc();
bx.curblock = bx.startblock;
bx.funcsym = s;
bx.scope_index = -1;
bx.classdec = cd;
bx.member = func;
bx.module = getModule();
irs.blx = &bx;
#if DMDV2
buildClosure(&irs);
#endif
#if 0
if (func->isSynchronized())
{
if (cd)
{ elem *esync;
if (func->isStatic())
{ // monitor is in ClassInfo
esync = el_ptr(cd->toSymbol());
}
else
{ // 'this' is the monitor
esync = el_var(sthis);
}
if (func->isStatic() || sbody->usesEH() ||
!(config.flags2 & CFG2seh))
{ // BUG: what if frequire or fensure uses EH?
sbody = new SynchronizedStatement(func->loc, esync, sbody);
}
else
{
#if TARGET_WINDOS
if (config.flags2 & CFG2seh)
{
/* The "jmonitor" uses an optimized exception handling frame
* which is a little shorter than the more general EH frame.
* It isn't strictly necessary.
*/
s->Sfunc->Fflags3 |= Fjmonitor;
}
#endif
el_free(esync);
}
}
else
{
error("synchronized function %s must be a member of a class", func->toChars());
}
}
#elif TARGET_WINDOS
if (func->isSynchronized() && cd && config.flags2 & CFG2seh &&
!func->isStatic() && !sbody->usesEH())
{
/* The "jmonitor" hack uses an optimized exception handling frame
* which is a little shorter than the more general EH frame.
*/
s->Sfunc->Fflags3 |= Fjmonitor;
}
#endif
sbody->toIR(&irs);
bx.curblock->BC = BCret;
f->Fstartblock = bx.startblock;
// einit = el_combine(einit,bx.init);
if (isCtorDeclaration())
{
assert(sthis);
for (b = f->Fstartblock; b; b = b->Bnext)
{
if (b->BC == BCret)
{
b->BC = BCretexp;
b->Belem = el_combine(b->Belem, el_var(sthis));
}
}
}
}
// If static constructor
#if DMDV2
if (isSharedStaticCtorDeclaration()) // must come first because it derives from StaticCtorDeclaration
{
elem *e = el_una(OPucall, TYvoid, el_var(s));
esharedctor = el_combine(esharedctor, e);
}
else
#endif
if (isStaticCtorDeclaration())
{
elem *e = el_una(OPucall, TYvoid, el_var(s));
ector = el_combine(ector, e);
}
// If static destructor
#if DMDV2
if (isSharedStaticDtorDeclaration()) // must come first because it derives from StaticDtorDeclaration
{
elem *e;
#if STATICCTOR
e = el_bin(OPcall, TYvoid, el_var(rtlsym[RTLSYM_FATEXIT]), el_ptr(s));
esharedctor = el_combine(esharedctor, e);
shareddtorcount++;
#else
SharedStaticDtorDeclaration *f = isSharedStaticDtorDeclaration();
assert(f);
if (f->vgate)
{ /* Increment destructor's vgate at construction time
*/
esharedctorgates.push(f);
}
e = el_una(OPucall, TYvoid, el_var(s));
eshareddtor = el_combine(e, eshareddtor);
#endif
}
else
#endif
if (isStaticDtorDeclaration())
{
elem *e;
#if STATICCTOR
e = el_bin(OPcall, TYvoid, el_var(rtlsym[RTLSYM_FATEXIT]), el_ptr(s));
ector = el_combine(ector, e);
dtorcount++;
#else
StaticDtorDeclaration *f = isStaticDtorDeclaration();
assert(f);
if (f->vgate)
{ /* Increment destructor's vgate at construction time
*/
ectorgates.push(f);
}
e = el_una(OPucall, TYvoid, el_var(s));
edtor = el_combine(e, edtor);
#endif
}
// If unit test
if (isUnitTestDeclaration())
{
elem *e = el_una(OPucall, TYvoid, el_var(s));
etest = el_combine(etest, e);
}
if (global.errors)
return;
writefunc(s);
if (isExport())
obj_export(s, Poffset);
for (size_t i = 0; i < irs.deferToObj->dim; i++)
{
Dsymbol *s = irs.deferToObj->tdata()[i];
s->toObjFile(0);
}
#if TARGET_LINUX || TARGET_OSX || TARGET_FREEBSD || TARGET_OPENBSD || TARGET_SOLARIS
// A hack to get a pointer to this function put in the .dtors segment
if (ident && memcmp(ident->toChars(), "_STD", 4) == 0)
obj_staticdtor(s);
#endif
#if DMDV2
if (irs.startaddress)
{
printf("Setting start address\n");
obj_startaddress(irs.startaddress);
}
#endif
}
char *template_mangle(symbol *s,param_t *arglist)
{
/* mangling ::= '$' template_name { type | expr }
type ::= "T" mangled type
expr ::= integer | string | address | float | double | long_double
integer ::= "I" dimension
string ::= "S" string
address ::= "R" zname
float ::= "F" hex_digits
double ::= "D" hex_digits
long_double ::= "L" hex_digits
*/
param_t *p;
assert(s);
symbol_debug(s);
//assert(s->Sclass == SCtemplate);
//printf("\ntemplate_mangle(s = '%s', arglist = %p)\n", s->Sident, arglist);
//arglist->print_list();
MangleInuse m;
mangle.znamei = 0;
mangle.argi = 0;
mangle.np = mangle.buf;
mangle.buf[BUFIDMAX + 1] = 0x55;
if (NEWTEMPMANGLE)
STR("?$");
else
CHAR('$');
// BUG: this is for templates nested inside class scopes.
// Need to check if it creates names that are properly unmanglable.
cpp_zname(s->Sident);
if (s->Sscope)
cpp_scope(s->Sscope);
for (p = arglist; p; p = p->Pnext)
{
if (p->Ptype)
{ /* Argument is a type */
if (!NEWTEMPMANGLE)
CHAR('T');
cpp_argument_list(p->Ptype, 1);
}
else if (p->Psym)
{
CHAR('V'); // this is a 'class' name, but it should be a 'template' name
cpp_ecsu_name(p->Psym);
}
else
{ /* Argument is an expression */
elem *e = p->Pelem;
tym_t ty = tybasic(e->ET->Tty);
char *p;
char a[2];
int ni;
char c;
L2:
switch (e->Eoper)
{ case OPconst:
switch (ty)
{ case TYfloat: ni = FLOATSIZE; c = 'F'; goto L1;
case TYdouble_alias:
case TYdouble: ni = DOUBLESIZE; c = 'D'; goto L1;
case TYldouble: ni = tysize(TYldouble); c = 'L'; goto L1;
L1:
if (NEWTEMPMANGLE)
CHAR('$');
CHAR(c);
p = (char *)&e->EV.Vdouble;
while (ni--)
{ char c;
#if __GNUC__
static char hex[16] =
{'0','1','2','3','4','5','6','7','8','9','A','B','C','D','E','F'};
#else
static char hex[16] = "0123456789ABCDEF";
#endif
c = *p++;
CHAR(hex[c & 15]);
CHAR(hex[(c >> 4) & 15]);
}
break;
default:
#ifdef DEBUG
if (!tyintegral(ty) && !tymptr(ty))
elem_print(e);
#endif
assert(tyintegral(ty) || tymptr(ty));
if (NEWTEMPMANGLE)
STR("$0");
else
CHAR('I');
cpp_dimension(el_tolongt(e));
break;
}
break;
case OPstring:
if (NEWTEMPMANGLE)
STR("$S");
else
CHAR('S');
if (e->EV.ss.Voffset)
synerr(EM_const_init); // constant initializer expected
cpp_string(e->EV.ss.Vstring,e->EV.ss.Vstrlen);
break;
case OPrelconst:
if (e->EV.sp.Voffset)
synerr(EM_const_init); // constant initializer expected
s = e->EV.sp.Vsym;
if (NEWTEMPMANGLE)
{ STR("$1");
cpp_decorated_name(s);
}
else
{ CHAR('R');
cpp_zname(s->Sident);
}
break;
case OPvar:
if (e->EV.sp.Vsym->Sflags & SFLvalue &&
tybasic(e->ET->Tty) != TYstruct)
{
e = e->EV.sp.Vsym->Svalue;
goto L2;
}
else if (e->EV.sp.Vsym->Sclass == SCconst /*&&
pstate.STintemplate*/)
{
CHAR('V'); // pretend to be a class name
cpp_zname(e->EV.sp.Vsym->Sident);
break;
}
default:
#if SCPP
#ifdef DEBUG
if (!errcnt)
elem_print(e);
#endif
synerr(EM_const_init); // constant initializer expected
assert(errcnt);
#endif
break;
}
}
}
*mangle.np = 0;
//printf("template_mangle() = '%s'\n", mangle.buf);
assert(strlen(mangle.buf) <= BUFIDMAX);
assert(mangle.buf[BUFIDMAX + 1] == 0x55);
return mangle.buf;
}
code *orthxmm(elem *e, regm_t *pretregs)
{
//printf("orthxmm(e = %p, *pretregs = %s)\n", e, regm_str(*pretregs));
elem *e1 = e->E1;
elem *e2 = e->E2;
// float + ifloat is not actually addition
if ((e->Eoper == OPadd || e->Eoper == OPmin) &&
((tyreal(e1->Ety) && tyimaginary(e2->Ety)) ||
(tyreal(e2->Ety) && tyimaginary(e1->Ety))))
{
regm_t retregs = *pretregs & XMMREGS;
if (!retregs)
retregs = XMMREGS;
unsigned reg;
regm_t rretregs;
unsigned rreg;
if (tyreal(e1->Ety))
{
reg = findreg(retregs);
rreg = findreg(retregs & ~mask[reg]);
retregs = mask[reg];
rretregs = mask[rreg];
}
else
{
// Pick the second register, not the first
rreg = findreg(retregs);
rretregs = mask[rreg];
reg = findreg(retregs & ~rretregs);
retregs = mask[reg];
}
assert(retregs && rretregs);
CodeBuilder cdb;
cdb.append(codelem(e1,&retregs,FALSE)); // eval left leaf
cdb.append(scodelem(e2, &rretregs, retregs, TRUE)); // eval right leaf
retregs |= rretregs;
if (e->Eoper == OPmin)
{
unsigned nretregs = XMMREGS & ~retregs;
unsigned sreg; // hold sign bit
unsigned sz = tysize(e1->Ety);
cdb.append(allocreg(&nretregs,&sreg,e2->Ety));
targ_size_t signbit = 0x80000000;
if (sz == 8)
signbit = 0x8000000000000000LL;
cdb.append(movxmmconst(sreg, sz, signbit, 0));
cdb.append(getregs(nretregs));
unsigned xop = (sz == 8) ? XORPD : XORPS; // XORPD/S rreg,sreg
cdb.gen2(xop,modregxrmx(3,rreg-XMM0,sreg-XMM0));
}
if (retregs != *pretregs)
cdb.append(fixresult(e,retregs,pretregs));
return cdb.finish();
}
regm_t retregs = *pretregs & XMMREGS;
if (!retregs)
retregs = XMMREGS;
CodeBuilder cdb;
cdb.append(codelem(e1,&retregs,FALSE)); // eval left leaf
unsigned reg = findreg(retregs);
regm_t rretregs = XMMREGS & ~retregs;
cdb.append(scodelem(e2, &rretregs, retregs, TRUE)); // eval right leaf
unsigned rreg = findreg(rretregs);
unsigned op = xmmoperator(e1->Ety, e->Eoper);
/* We should take advantage of mem addressing modes for OP XMM,MEM
* but we do not at the moment.
*/
if (OTrel(e->Eoper))
{
retregs = mPSW;
cdb.gen2(op,modregxrmx(3,rreg-XMM0,reg-XMM0));
checkSetVex(cdb.last(), e1->Ety);
return cdb.finish();
}
else
cdb.append(getregs(retregs));
cdb.gen2(op,modregxrmx(3,reg-XMM0,rreg-XMM0));
checkSetVex(cdb.last(), e1->Ety);
if (retregs != *pretregs)
cdb.append(fixresult(e,retregs,pretregs));
return cdb.finish();
}
code *xmmeq(elem *e, unsigned op, elem *e1, elem *e2,regm_t *pretregs)
{
tym_t tymll;
unsigned reg;
int i;
code cs;
elem *e11;
bool regvar; /* TRUE means evaluate into register variable */
regm_t varregm;
unsigned varreg;
targ_int postinc;
//printf("xmmeq(e1 = %p, e2 = %p, *pretregs = %s)\n", e1, e2, regm_str(*pretregs));
int e2oper = e2->Eoper;
tym_t tyml = tybasic(e1->Ety); /* type of lvalue */
regm_t retregs = *pretregs;
if (!(retregs & XMMREGS))
retregs = XMMREGS; // pick any XMM reg
bool aligned = xmmIsAligned(e1);
cs.Iop = (op == OPeq) ? xmmstore(tyml, aligned) : op;
regvar = FALSE;
varregm = 0;
if (config.flags4 & CFG4optimized)
{
// Be careful of cases like (x = x+x+x). We cannot evaluate in
// x if x is in a register.
if (isregvar(e1,&varregm,&varreg) && // if lvalue is register variable
doinreg(e1->EV.sp.Vsym,e2) && // and we can compute directly into it
varregm & XMMREGS
)
{ regvar = TRUE;
retregs = varregm;
reg = varreg; /* evaluate directly in target register */
}
}
if (*pretregs & mPSW && !EOP(e1)) // if evaluating e1 couldn't change flags
{ // Be careful that this lines up with jmpopcode()
retregs |= mPSW;
*pretregs &= ~mPSW;
}
CodeBuilder cdb;
cdb.append(scodelem(e2,&retregs,0,TRUE)); // get rvalue
// Look for special case of (*p++ = ...), where p is a register variable
if (e1->Eoper == OPind &&
((e11 = e1->E1)->Eoper == OPpostinc || e11->Eoper == OPpostdec) &&
e11->E1->Eoper == OPvar &&
e11->E1->EV.sp.Vsym->Sfl == FLreg
)
{
postinc = e11->E2->EV.Vint;
if (e11->Eoper == OPpostdec)
postinc = -postinc;
cdb.append(getlvalue(&cs,e11,RMstore | retregs));
freenode(e11->E2);
}
else
{ postinc = 0;
cdb.append(getlvalue(&cs,e1,RMstore | retregs)); // get lvalue (cl == CNIL if regvar)
}
cdb.append(getregs_imm(regvar ? varregm : 0));
reg = findreg(retregs & XMMREGS);
cs.Irm |= modregrm(0,(reg - XMM0) & 7,0);
if ((reg - XMM0) & 8)
cs.Irex |= REX_R;
// Do not generate mov from register onto itself
if (!(regvar && reg == XMM0 + ((cs.Irm & 7) | (cs.Irex & REX_B ? 8 : 0))))
{
cdb.gen(&cs); // MOV EA+offset,reg
if (op == OPeq)
checkSetVex(cdb.last(), tyml);
}
if (e1->Ecount || // if lvalue is a CSE or
regvar) // rvalue can't be a CSE
{
cdb.append(getregs_imm(retregs)); // necessary if both lvalue and
// rvalue are CSEs (since a reg
// can hold only one e at a time)
cssave(e1,retregs,EOP(e1)); // if lvalue is a CSE
}
cdb.append(fixresult(e,retregs,pretregs));
Lp:
if (postinc)
{
int reg = findreg(idxregm(&cs));
if (*pretregs & mPSW)
{ // Use LEA to avoid touching the flags
unsigned rm = cs.Irm & 7;
if (cs.Irex & REX_B)
rm |= 8;
cdb.genc1(0x8D,buildModregrm(2,reg,rm),FLconst,postinc);
if (tysize(e11->E1->Ety) == 8)
code_orrex(cdb.last(), REX_W);
}
else if (I64)
{
cdb.genc2(0x81,modregrmx(3,0,reg),postinc);
if (tysize(e11->E1->Ety) == 8)
code_orrex(cdb.last(), REX_W);
}
else
{
if (postinc == 1)
cdb.gen1(0x40 + reg); // INC reg
else if (postinc == -(targ_int)1)
cdb.gen1(0x48 + reg); // DEC reg
else
{
cdb.genc2(0x81,modregrm(3,0,reg),postinc);
}
}
}
freenode(e1);
return cdb.finish();
}
void checkSetVex(code *c, tym_t ty)
{
if (config.avx || tysize(ty) == 32)
{
unsigned vreg = (c->Irm >> 3) & 7;
if (c->Irex & REX_R)
vreg |= 8;
int VEX_L = 0;
// TODO: This is too simplistic, depending on the instruction, vex.vvvv
// encodes NDS, NDD, DDS, or no operand (NOO). The code below assumes
// NDS (non-destructive source), except for the incomplete list of 2
// operand instructions (NOO) handled by the switch.
switch (c->Iop)
{
case LODSS:
case LODSD:
case STOSS:
case STOSD:
if ((c->Irm & 0xC0) == 0xC0)
break;
case LODAPS:
case LODUPS:
case LODAPD:
case LODUPD:
case LODDQA:
case LODDQU:
case LODD:
case LODQ:
case STOAPS:
case STOUPS:
case STOAPD:
case STOUPD:
case STODQA:
case STODQU:
case STOD:
case STOQ:
case COMISS:
case COMISD:
case UCOMISS:
case UCOMISD:
case MOVDDUP:
case MOVSHDUP:
case MOVSLDUP:
case VBROADCASTSS:
case PSHUFD:
case PSHUFHW:
case PSHUFLW:
vreg = 0; // for 2 operand vex instructions
break;
case VBROADCASTSD:
case VBROADCASTF128:
VEX_L = 1;
vreg = 0; // for 2 operand vex instructions
break;
}
unsigned op = 0xC4000000 | (c->Iop & 0xFF);
switch (c->Iop & 0xFFFFFF00)
{
#define MM_PP(mm, pp) ((mm << 16) | (pp << 8))
case 0x00000F00: op |= MM_PP(1,0); break;
case 0x00660F00: op |= MM_PP(1,1); break;
case 0x00F30F00: op |= MM_PP(1,2); break;
case 0x00F20F00: op |= MM_PP(1,3); break;
case 0x660F3800: op |= MM_PP(2,1); break;
case 0x660F3A00: op |= MM_PP(3,1); break;
default:
printf("Iop = %x\n", c->Iop);
assert(0);
#undef MM_PP
}
c->Iop = op;
c->Ivex.pfx = 0xC4;
c->Ivex.r = !(c->Irex & REX_R);
c->Ivex.x = !(c->Irex & REX_X);
c->Ivex.b = !(c->Irex & REX_B);
c->Ivex.w = (c->Irex & REX_W) != 0;
c->Ivex.l = (tysize(ty) == 32) | VEX_L;
c->Ivex.vvvv = ~vreg;
c->Iflags |= CFvex;
checkSetVex3(c);
}
/***************
* Generate code for OPvecfill (broadcast).
* OPvecfill takes the single value in e1 and
* fills the vector type with it.
*/
code *cdvecfill(elem *e, regm_t *pretregs)
{
//printf("cdvecfill(e = %p, *pretregs = %s)\n",e,regm_str(*pretregs));
regm_t retregs = *pretregs & XMMREGS;
if (!retregs)
retregs = XMMREGS;
CodeBuilder cdb;
code *c;
code cs;
elem *e1 = e->E1;
#if 0
if ((e1->Eoper == OPind && !e1->Ecount) || e1->Eoper == OPvar)
{
cr = getlvalue(&cs, e1, RMload | retregs); // get addressing mode
}
else
{
unsigned rretregs = XMMREGS & ~retregs;
cr = scodelem(op2, &rretregs, retregs, TRUE);
unsigned rreg = findreg(rretregs) - XMM0;
cs.Irm = modregrm(3,0,rreg & 7);
cs.Iflags = 0;
cs.Irex = 0;
if (rreg & 8)
cs.Irex |= REX_B;
}
#endif
unsigned reg;
unsigned rreg;
unsigned varreg;
regm_t varregm;
tym_t ty = tybasic(e->Ety);
switch (ty)
{
case TYfloat4:
case TYfloat8:
if (config.avx &&
((e1->Eoper == OPind && !e1->Ecount) || e1->Eoper == OPvar && !isregvar(e1,&varregm,&varreg)) ||
tysize(ty) == 32 && !isregvar(e1,&varregm,&varreg)
)
{
Lint:
if (e1->Eoper == OPvar)
e1->EV.sp.Vsym->Sflags &= ~GTregcand;
// VBROADCASTSS XMM,MEM
cdb.append(getlvalue(&cs, e1, 0)); // get addressing mode
assert((cs.Irm & 0xC0) != 0xC0); // AVX1 doesn't have register source operands
cdb.append(allocreg(&retregs,®,ty));
cs.Iop = VBROADCASTSS;
cs.Irex &= ~REX_W;
code_newreg(&cs,reg - XMM0);
checkSetVex(&cs,ty);
cdb.gen(&cs);
}
else
{
// SHUFPS XMM0,XMM0,0 0F C6 /r ib
c = codelem(e1,&retregs,FALSE); // eval left leaf
cdb.append(c);
reg = findreg(retregs) - XMM0;
cdb.append(getregs(retregs));
cs.Iop = SHUFPS;
cs.Irm = modregxrmx(3,reg,reg);
cs.Iflags = 0;
cs.IFL2 = FLconst;
cs.IEV2.Vsize_t = 0;
if (config.avx >= 2 || tysize(ty) == 32)
{
// VBROADCASTSS XMM,XMM
cs.Iop = VBROADCASTSS;
checkSetVex(&cs, ty);
}
cdb.gen(&cs);
}
break;
case TYdouble2:
case TYdouble4:
if (config.avx &&
((e1->Eoper == OPind && !e1->Ecount) || e1->Eoper == OPvar && !isregvar(e1,&varregm,&varreg)) ||
tysize(ty) == 32 && !isregvar(e1,&varregm,&varreg)
)
{
if (e1->Eoper == OPvar)
e1->EV.sp.Vsym->Sflags &= ~GTregcand;
// VBROADCASTSD XMM,MEM
cdb.append(getlvalue(&cs, e1, 0)); // get addressing mode
assert((cs.Irm & 0xC0) != 0xC0); // AVX1 doesn't have register source operands
cdb.append(allocreg(&retregs,®,ty));
cs.Iop = VBROADCASTSD;
cs.Irex &= ~REX_W;
code_newreg(&cs,reg - XMM0);
checkSetVex(&cs,ty);
cdb.gen(&cs);
}
else
{
// UNPCKLPD XMM0,XMM0 66 0F 14 /r
c = codelem(e1,&retregs,FALSE); // eval left leaf
cdb.append(c);
reg = findreg(retregs) - XMM0;
cdb.append(getregs(retregs));
cs.Iop = UNPCKLPD;
cs.Irm = modregxrmx(3,reg,reg);
cs.Iflags = 0;
if (config.avx >= 2 || tysize(ty) == 32)
{
// VBROADCASTSD XMM,XMM
cs.Iop = VBROADCASTSD;
checkSetVex(&cs, ty);
}
cdb.gen(&cs);
}
break;
case TYschar16:
case TYuchar16:
case TYschar32:
case TYuchar32:
{
/* MOVD XMM0,r
* PUNPCKLBW XMM0,XMM0
* PUNPCKLWD XMM0,XMM0
* PSHUFD XMM0,XMM0,0
*/
regm_t regm = ALLREGS;
c = codelem(e1,®m,FALSE); // eval left leaf
cdb.append(c);
unsigned r = findreg(regm);
c = allocreg(&retregs,®, e->Ety);
cdb.append(c);
reg -= XMM0;
cdb.gen2(LODD,modregxrmx(3,reg,r)); // MOVD reg,r
checkSetVex(cdb.last(),TYschar16);
cs.Iop = PUNPCKLBW;
cs.Irm = modregxrmx(3,reg,reg);
cs.Iflags = 0;
cdb.gen(&cs);
cs.Iop = PUNPCKLWD;
cdb.gen(&cs);
cs.Iop = PSHUFD;
cs.IFL2 = FLconst;
cs.IEV2.Vsize_t = 0;
checkSetVex(&cs,TYschar16);
cdb.gen(&cs);
if (tysize(ty) == 32)
{
// VINSERTF128 YMM0,YMM0,XMM0,1
cs.Iop = VINSERTF128;
cs.Irm = modregxrmx(3,reg,reg);
cs.Iflags = 0;
cs.IFL2 = FLconst;
cs.IEV2.Vsize_t = 1;
checkSetVex(&cs,ty);
cdb.gen(&cs);
}
break;
}
case TYshort8:
case TYushort8:
case TYshort16:
case TYushort16:
{
regm_t regm = ALLREGS;
c = codelem(e1,®m,FALSE); // eval left leaf
cdb.append(c);
unsigned r = findreg(regm);
if (config.avx || tysize(ty) == 32)
{
/*
* VPXOR XMM0,XMM0,XMM0
* VPINSRW XMM0,XMM0,r,0
* VPINSRW XMM0,XMM0,r,1
* VPINSRW XMM0,XMM0,r,2
* VPINSRW XMM0,XMM0,r,3
*/
cdb.append(allocreg(&retregs,®, ty));
cdb.gen2(PXOR,modregxrmx(3,reg-XMM0,reg-XMM0));
checkSetVex(cdb.last(), TYshort8);
for (int i = 0; i < tysize(ty) / 4; ++i)
{
cdb.genc2(PINSRW,modregxrmx(3,reg-XMM0,r),i);
checkSetVex(cdb.last(), TYshort8);
}
if (tysize(ty) == 32)
{
// VINSERTF128 YMM0,YMM0,XMM0,1
cs.Iop = VINSERTF128;
cs.Irm = modregxrmx(3,reg-XMM0,reg-XMM0);
cs.Iflags = 0;
cs.IFL2 = FLconst;
cs.IEV2.Vsize_t = 1;
checkSetVex(&cs,ty);
cdb.gen(&cs);
}
else
{
// VPSHUFD XMM0,XMM0,0
cs.Iop = PSHUFD;
cs.Irm = modregxrmx(3,reg-XMM0,reg-XMM0);
cs.Iflags = 0;
cs.IFL2 = FLconst;
cs.IEV2.Vsize_t = 0;
checkSetVex(&cs,ty);
cdb.gen(&cs);
}
}
else
{
/* MOVD XMM0,r
* PUNPCKLWD XMM0,XMM0
* PSHUFD XMM0,XMM0,0
*/
c = allocreg(&retregs,®, e->Ety);
cdb.append(c);
reg -= XMM0;
cdb.gen2(LODD,modregxrmx(3,reg,r)); // MOVD reg,r
checkSetVex(cdb.last(),e->Ety);
cs.Iop = PUNPCKLWD;
cs.Irm = modregxrmx(3,reg,reg);
cs.Iflags = 0;
cdb.gen(&cs);
cs.Iop = PSHUFD;
cs.IFL2 = FLconst;
cs.IEV2.Vsize_t = 0;
cdb.gen(&cs);
}
break;
}
case TYlong8:
case TYulong8:
case TYlong4:
case TYulong4:
{
if (config.avx &&
((e1->Eoper == OPind && !e1->Ecount) || e1->Eoper == OPvar && !isregvar(e1,&varregm,&varreg)) ||
tysize(ty) == 32 && !isregvar(e1,&varregm,&varreg))
{
goto Lint;
}
/* MOVD XMM1,r
* PSHUFD XMM0,XMM1,0
*/
regm_t regm = ALLREGS;
c = codelem(e1,®m,FALSE); // eval left leaf
cdb.append(c);
unsigned r = findreg(regm);
c = allocreg(&retregs,®, e->Ety);
cdb.append(c);
reg -= XMM0;
cdb.gen2(LODD,modregxrmx(3,reg,r)); // MOVD reg,r
cs.Iop = PSHUFD;
cs.Irm = modregxrmx(3,reg,reg);
cs.Iflags = 0;
cs.IFL2 = FLconst;
cs.IEV2.Vsize_t = 0;
if (config.avx >= 2 || tysize(ty) == 32)
{
// VBROADCASTSS XMM,XMM
cs.Iop = VBROADCASTSS;
checkSetVex(&cs, ty);
}
cdb.gen(&cs);
break;
}
case TYllong2:
case TYullong2:
case TYllong4:
case TYullong4:
if (config.avx || tysize(ty) >= 32)
{
if (e1->Eoper == OPvar)
e1->EV.sp.Vsym->Sflags &= ~GTregcand;
// VMOVDDUP XMM,MEM
cdb.append(getlvalue(&cs, e1, 0)); // get addressing mode
if ((cs.Irm & 0xC0) == 0xC0)
{
unsigned sreg = ((cs.Irm & 7) | (cs.Irex & REX_B ? 8 : 0));
regm_t sregm = XMMREGS;
cdb.append(fixresult(e1, mask[sreg], &sregm));
unsigned rmreg = findreg(sregm);
cs.Irm = (cs.Irm & ~7) | ((rmreg - XMM0) & 7);
if ((rmreg - XMM0) & 8)
cs.Irex |= REX_B;
else
cs.Irex &= ~REX_B;
}
cdb.append(allocreg(&retregs,®,ty));
if (config.avx >= 2 || tysize(ty) >= 32)
{
cs.Iop = VBROADCASTSD;
cs.Irex &= ~REX_W;
}
else
cs.Iop = MOVDDUP;
code_newreg(&cs,reg - XMM0);
checkSetVex(&cs,ty);
cdb.gen(&cs);
}
else
{
/* MOVQ XMM0,mem128
* PUNPCKLQDQ XMM0,XMM0
*/
c = codelem(e1,&retregs,FALSE); // eval left leaf
cdb.append(c);
unsigned reg = findreg(retregs);
reg -= XMM0;
//cdb.gen2(LODD,modregxrmx(3,reg,r)); // MOVQ reg,r
cs.Iop = PUNPCKLQDQ;
cs.Irm = modregxrmx(3,reg,reg);
cs.Iflags = 0;
cdb.gen(&cs);
}
break;
default:
assert(0);
}
c = fixresult(e,retregs,pretregs);
cdb.append(c);
return cdb.finish();
}
static void sliceStructs_Gather(SymInfo *sia, elem *e)
{
while (1)
{
switch (e->Eoper)
{
case OPvar:
{
SYMIDX si = e->EV.sp.Vsym->Ssymnum;
if (si >= 0 && sia[si].canSlice)
{
assert(si < globsym.top);
unsigned sz = tysize(e->Ety);
if (sz == 2 * REGSIZE && !tyfv(e->Ety))
{
// Rewritten as OPpair later
}
else if (sz == REGSIZE &&
(e->Eoffset == 0 || e->Eoffset == REGSIZE))
{
if (!sia[si].accessSlice)
{
sia[si].ty0 = TYnptr;
sia[si].ty1 = TYnptr;
}
sia[si].accessSlice = true;
if (e->Eoffset == 0)
sia[si].ty0 = tybasic(e->Ety);
else
sia[si].ty1 = tybasic(e->Ety);
}
else
{
sia[si].canSlice = false;
}
}
return;
}
default:
if (OTassign(e->Eoper))
{
if (OTbinary(e->Eoper))
sliceStructs_Gather(sia, e->E2);
// Assignment to a whole var will disallow SROA
if (e->E1->Eoper == OPvar)
{
elem *e1 = e->E1;
SYMIDX si = e1->EV.sp.Vsym->Ssymnum;
if (si >= 0 && sia[si].canSlice)
{
assert(si < globsym.top);
if (tysize(e1->Ety) != REGSIZE ||
(e1->Eoffset != 0 && e1->Eoffset != REGSIZE))
{
sia[si].canSlice = false;
}
}
return;
}
e = e->E1;
break;
}
if (OTunary(e->Eoper))
{
e = e->E1;
break;
}
if (OTbinary(e->Eoper))
{
sliceStructs_Gather(sia, e->E2);
e = e->E1;
break;
}
return;
}
}
}
static void sliceStructs_Gather(SymInfo *sia, elem *e)
{
while (1)
{
switch (e->Eoper)
{
case OPvar:
{
SYMIDX si = e->EV.sp.Vsym->Ssymnum;
if (si >= 0 && sia[si].canSlice)
{
assert(si < globsym.top);
unsigned sz = tysize(e->Ety);
if (sz == 2 * REGSIZE)
{
// Rewrite as OPpair later
sia[si].usePair = true;
/* OPpair cannot handle XMM registers, cdpair() and fixresult()
*/
if (tyfloating(sia[si].ty0) || tyfloating(sia[si].ty1))
sia[si].canSlice = false;
}
else if (sz == REGSIZE &&
(e->Eoffset == 0 || e->Eoffset == REGSIZE))
{
if (!sia[si].accessSlice)
{
sia[si].ty0 = TYnptr;
sia[si].ty1 = TYnptr;
}
sia[si].accessSlice = true;
if (e->Eoffset == 0)
sia[si].ty0 = tybasic(e->Ety);
else
sia[si].ty1 = tybasic(e->Ety);
// Cannot slice float fields if the symbol is also accessed using OPpair (see above)
if (sia[si].usePair && (tyfloating(sia[si].ty0) || tyfloating(sia[si].ty1)))
sia[si].canSlice = false;
}
else
{
sia[si].canSlice = false;
}
}
return;
}
default:
if (OTassign(e->Eoper))
{
if (OTbinary(e->Eoper))
sliceStructs_Gather(sia, e->E2);
// Assignment to a whole var will disallow SROA
if (e->E1->Eoper == OPvar)
{
elem *e1 = e->E1;
SYMIDX si = e1->EV.sp.Vsym->Ssymnum;
if (si >= 0 && sia[si].canSlice)
{
assert(si < globsym.top);
if (tysize(e1->Ety) != REGSIZE ||
(e1->Eoffset != 0 && e1->Eoffset != REGSIZE))
{
sia[si].canSlice = false;
}
}
return;
}
e = e->E1;
break;
}
if (OTunary(e->Eoper))
{
e = e->E1;
break;
}
if (OTbinary(e->Eoper))
{
sliceStructs_Gather(sia, e->E2);
e = e->E1;
break;
}
return;
}
}
}
