void
dostkoff(void)
{
Prog *p, *q, *q1;
int32 autoffset, deltasp;
int a, pcsize;
uint32 moreconst1, moreconst2, i;
for(i=0; i<nelem(morename); i++) {
symmorestack[i] = lookup(morename[i], 0);
if(symmorestack[i]->type != STEXT)
diag("morestack trampoline not defined - %s", morename[i]);
pmorestack[i] = symmorestack[i]->text;
}
for(cursym = textp; cursym != nil; cursym = cursym->next) {
if(cursym->text == nil || cursym->text->link == nil)
continue;
p = cursym->text;
parsetextconst(p->to.offset);
autoffset = textstksiz;
if(autoffset < 0)
autoffset = 0;
q = P;
if((p->from.scale & NOSPLIT) && autoffset >= StackSmall)
diag("nosplit func likely to overflow stack");
if(!(p->from.scale & NOSPLIT)) {
p = appendp(p); // load g into CX
p->as = AMOVQ;
if(HEADTYPE == Hlinux || HEADTYPE == Hfreebsd
|| HEADTYPE == Hopenbsd || HEADTYPE == Hnetbsd) // ELF uses FS
p->from.type = D_INDIR+D_FS;
else
p->from.type = D_INDIR+D_GS;
p->from.offset = tlsoffset+0;
p->to.type = D_CX;
if(HEADTYPE == Hwindows) {
// movq %gs:0x28, %rcx
// movq (%rcx), %rcx
p->as = AMOVQ;
p->from.type = D_INDIR+D_GS;
p->from.offset = 0x28;
p->to.type = D_CX;
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_INDIR+D_CX;
p->from.offset = 0;
p->to.type = D_CX;
}
if(debug['K']) {
// 6l -K means check not only for stack
// overflow but stack underflow.
// On underflow, INT 3 (breakpoint).
// Underflow itself is rare but this also
// catches out-of-sync stack guard info
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_INDIR+D_CX;
p->from.offset = 8;
p->to.type = D_SP;
p = appendp(p);
p->as = AJHI;
p->to.type = D_BRANCH;
p->to.offset = 4;
q1 = p;
p = appendp(p);
p->as = AINT;
p->from.type = D_CONST;
p->from.offset = 3;
p = appendp(p);
p->as = ANOP;
q1->pcond = p;
}
if(autoffset < StackBig) { // do we need to call morestack?
if(autoffset <= StackSmall) {
// small stack
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_SP;
p->to.type = D_INDIR+D_CX;
} else {
// large stack
p = appendp(p);
p->as = ALEAQ;
p->from.type = D_INDIR+D_SP;
p->from.offset = -(autoffset-StackSmall);
p->to.type = D_AX;
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_AX;
p->to.type = D_INDIR+D_CX;
}
// common
p = appendp(p);
p->as = AJHI;
p->to.type = D_BRANCH;
p->to.offset = 4;
q = p;
}
/* 160 comes from 3 calls (3*8) 4 safes (4*8) and 104 guard */
moreconst1 = 0;
if(autoffset+160+textarg > 4096)
moreconst1 = (autoffset+160) & ~7LL;
moreconst2 = textarg;
// 4 varieties varieties (const1==0 cross const2==0)
// and 6 subvarieties of (const1==0 and const2!=0)
p = appendp(p);
if(moreconst1 == 0 && moreconst2 == 0) {
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[0];
p->to.sym = symmorestack[0];
} else
if(moreconst1 != 0 && moreconst2 == 0) {
p->as = AMOVL;
p->from.type = D_CONST;
p->from.offset = moreconst1;
p->to.type = D_AX;
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[1];
p->to.sym = symmorestack[1];
} else
if(moreconst1 == 0 && moreconst2 <= 48 && moreconst2%8 == 0) {
i = moreconst2/8 + 3;
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[i];
p->to.sym = symmorestack[i];
} else
if(moreconst1 == 0 && moreconst2 != 0) {
p->as = AMOVL;
p->from.type = D_CONST;
p->from.offset = moreconst2;
p->to.type = D_AX;
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[2];
p->to.sym = symmorestack[2];
} else {
p->as = AMOVQ;
p->from.type = D_CONST;
p->from.offset = (uint64)moreconst2 << 32;
p->from.offset |= moreconst1;
p->to.type = D_AX;
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[3];
p->to.sym = symmorestack[3];
}
}
if(q != P)
q->pcond = p->link;
if(autoffset) {
p = appendp(p);
p->as = AADJSP;
p->from.type = D_CONST;
p->from.offset = autoffset;
p->spadj = autoffset;
if(q != P)
q->pcond = p;
}
deltasp = autoffset;
if(debug['K'] > 1 && autoffset) {
// 6l -KK means double-check for stack overflow
// even after calling morestack and even if the
// function is marked as nosplit.
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_INDIR+D_CX;
p->from.offset = 0;
p->to.type = D_BX;
p = appendp(p);
p->as = ASUBQ;
p->from.type = D_CONST;
p->from.offset = StackSmall+32;
p->to.type = D_BX;
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_SP;
p->to.type = D_BX;
p = appendp(p);
p->as = AJHI;
p->to.type = D_BRANCH;
q1 = p;
p = appendp(p);
p->as = AINT;
p->from.type = D_CONST;
p->from.offset = 3;
p = appendp(p);
p->as = ANOP;
q1->pcond = p;
}
for(; p != P; p = p->link) {
pcsize = p->mode/8;
a = p->from.type;
if(a == D_AUTO)
p->from.offset += deltasp;
if(a == D_PARAM)
p->from.offset += deltasp + pcsize;
a = p->to.type;
if(a == D_AUTO)
p->to.offset += deltasp;
if(a == D_PARAM)
p->to.offset += deltasp + pcsize;
switch(p->as) {
default:
continue;
case APUSHL:
case APUSHFL:
deltasp += 4;
p->spadj = 4;
continue;
case APUSHQ:
case APUSHFQ:
deltasp += 8;
p->spadj = 8;
continue;
case APUSHW:
case APUSHFW:
deltasp += 2;
p->spadj = 2;
continue;
case APOPL:
case APOPFL:
deltasp -= 4;
p->spadj = -4;
continue;
case APOPQ:
case APOPFQ:
deltasp -= 8;
p->spadj = -8;
continue;
case APOPW:
case APOPFW:
deltasp -= 2;
p->spadj = -2;
continue;
case ARET:
break;
}
if(autoffset != deltasp)
diag("unbalanced PUSH/POP");
if(autoffset) {
p->as = AADJSP;
p->from.type = D_CONST;
p->from.offset = -autoffset;
p->spadj = -autoffset;
p = appendp(p);
p->as = ARET;
// If there are instructions following
// this ARET, they come from a branch
// with the same stackframe, so undo
// the cleanup.
p->spadj = +autoffset;
}
}
}
}
void
dostkoff(void)
{
Prog *p, *q, *q1;
int32 autoffset, deltasp;
int a, pcsize;
uint32 moreconst1, moreconst2, i;
Sym *gmsym;
gmsym = lookup("runtime.tlsgm", 0);
for(i=0; i<nelem(morename); i++) {
symmorestack[i] = lookup(morename[i], 0);
if(symmorestack[i]->type != STEXT)
diag("morestack trampoline not defined - %s", morename[i]);
pmorestack[i] = symmorestack[i]->text;
}
for(cursym = textp; cursym != nil; cursym = cursym->next) {
if(cursym->text == nil || cursym->text->link == nil)
continue;
p = cursym->text;
parsetextconst(p->to.offset);
autoffset = textstksiz;
if(autoffset < 0)
autoffset = 0;
if(autoffset < StackSmall && !(p->from.scale & NOSPLIT)) {
for(q = p; q != P; q = q->link)
if(q->as == ACALL)
goto noleaf;
p->from.scale |= NOSPLIT;
noleaf:;
}
q = P;
q1 = P;
if((p->from.scale & NOSPLIT) && autoffset >= StackSmall)
diag("nosplit func likely to overflow stack");
if(!(p->from.scale & NOSPLIT)) {
if(flag_shared) {
// Load TLS offset with MOVQ $runtime.tlsgm(SB), CX
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_EXTERN;
p->from.sym = gmsym;
p->to.type = D_CX;
}
p = appendp(p); // load g into CX
p->as = AMOVQ;
if(HEADTYPE == Hlinux || HEADTYPE == Hfreebsd
|| HEADTYPE == Hopenbsd || HEADTYPE == Hnetbsd
|| HEADTYPE == Hplan9x64 || HEADTYPE == Hdragonfly)
// ELF uses FS
p->from.type = D_INDIR+D_FS;
else
p->from.type = D_INDIR+D_GS;
if(flag_shared) {
// Add TLS offset stored in CX
p->from.index = p->from.type - D_INDIR;
p->from.type = D_INDIR + D_CX;
}
p->from.offset = tlsoffset+0;
p->to.type = D_CX;
if(HEADTYPE == Hwindows) {
// movq %gs:0x28, %rcx
// movq (%rcx), %rcx
p->as = AMOVQ;
p->from.type = D_INDIR+D_GS;
p->from.offset = 0x28;
p->to.type = D_CX;
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_INDIR+D_CX;
p->from.offset = 0;
p->to.type = D_CX;
}
if(debug['K']) {
// 6l -K means check not only for stack
// overflow but stack underflow.
// On underflow, INT 3 (breakpoint).
// Underflow itself is rare but this also
// catches out-of-sync stack guard info
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_INDIR+D_CX;
p->from.offset = 8;
p->to.type = D_SP;
p = appendp(p);
p->as = AJHI;
p->to.type = D_BRANCH;
p->to.offset = 4;
q1 = p;
p = appendp(p);
p->as = AINT;
p->from.type = D_CONST;
p->from.offset = 3;
p = appendp(p);
p->as = ANOP;
q1->pcond = p;
}
q1 = P;
if(autoffset <= StackSmall) {
// small stack: SP <= stackguard
// CMPQ SP, stackguard
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_SP;
p->to.type = D_INDIR+D_CX;
} else if(autoffset <= StackBig) {
// large stack: SP-framesize <= stackguard-StackSmall
// LEAQ -xxx(SP), AX
// CMPQ AX, stackguard
p = appendp(p);
p->as = ALEAQ;
p->from.type = D_INDIR+D_SP;
p->from.offset = -(autoffset-StackSmall);
p->to.type = D_AX;
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_AX;
p->to.type = D_INDIR+D_CX;
} else {
// Such a large stack we need to protect against wraparound.
// If SP is close to zero:
// SP-stackguard+StackGuard <= framesize + (StackGuard-StackSmall)
// The +StackGuard on both sides is required to keep the left side positive:
// SP is allowed to be slightly below stackguard. See stack.h.
//
// Preemption sets stackguard to StackPreempt, a very large value.
// That breaks the math above, so we have to check for that explicitly.
// MOVQ stackguard, CX
// CMPQ CX, $StackPreempt
// JEQ label-of-call-to-morestack
// LEAQ StackGuard(SP), AX
// SUBQ CX, AX
// CMPQ AX, $(autoffset+(StackGuard-StackSmall))
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_INDIR+D_CX;
p->from.offset = 0;
p->to.type = D_SI;
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_SI;
p->to.type = D_CONST;
p->to.offset = StackPreempt;
p = appendp(p);
p->as = AJEQ;
p->to.type = D_BRANCH;
q1 = p;
p = appendp(p);
p->as = ALEAQ;
p->from.type = D_INDIR+D_SP;
p->from.offset = StackGuard;
p->to.type = D_AX;
p = appendp(p);
p->as = ASUBQ;
p->from.type = D_SI;
p->to.type = D_AX;
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_AX;
p->to.type = D_CONST;
p->to.offset = autoffset+(StackGuard-StackSmall);
}
// common
p = appendp(p);
p->as = AJHI;
p->to.type = D_BRANCH;
q = p;
// If we ask for more stack, we'll get a minimum of StackMin bytes.
// We need a stack frame large enough to hold the top-of-stack data,
// the function arguments+results, our caller's PC, our frame,
// a word for the return PC of the next call, and then the StackLimit bytes
// that must be available on entry to any function called from a function
// that did a stack check. If StackMin is enough, don't ask for a specific
// amount: then we can use the custom functions and save a few
// instructions.
moreconst1 = 0;
if(StackTop + textarg + PtrSize + autoffset + PtrSize + StackLimit >= StackMin)
moreconst1 = autoffset;
moreconst2 = textarg;
if(moreconst2 == 1) // special marker
moreconst2 = 0;
if((moreconst2&7) != 0)
diag("misaligned argument size in stack split");
// 4 varieties varieties (const1==0 cross const2==0)
// and 6 subvarieties of (const1==0 and const2!=0)
p = appendp(p);
if(moreconst1 == 0 && moreconst2 == 0) {
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[0];
p->to.sym = symmorestack[0];
} else
if(moreconst1 != 0 && moreconst2 == 0) {
p->as = AMOVL;
p->from.type = D_CONST;
p->from.offset = moreconst1;
p->to.type = D_AX;
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[1];
p->to.sym = symmorestack[1];
} else
if(moreconst1 == 0 && moreconst2 <= 48 && moreconst2%8 == 0) {
i = moreconst2/8 + 3;
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[i];
p->to.sym = symmorestack[i];
} else
if(moreconst1 == 0 && moreconst2 != 0) {
p->as = AMOVL;
p->from.type = D_CONST;
p->from.offset = moreconst2;
p->to.type = D_AX;
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[2];
p->to.sym = symmorestack[2];
} else {
p->as = AMOVQ;
p->from.type = D_CONST;
p->from.offset = (uint64)moreconst2 << 32;
p->from.offset |= moreconst1;
p->to.type = D_AX;
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[3];
p->to.sym = symmorestack[3];
}
p = appendp(p);
p->as = AJMP;
p->to.type = D_BRANCH;
p->pcond = cursym->text->link;
}
if(q != P)
q->pcond = p->link;
if(q1 != P)
q1->pcond = q->link;
if(autoffset) {
p = appendp(p);
p->as = AADJSP;
p->from.type = D_CONST;
p->from.offset = autoffset;
p->spadj = autoffset;
if(q != P)
q->pcond = p;
} else {
// zero-byte stack adjustment.
// Insert a fake non-zero adjustment so that stkcheck can
// recognize the end of the stack-splitting prolog.
p = appendp(p);
p->as = ANOP;
p->spadj = -PtrSize;
p = appendp(p);
p->as = ANOP;
p->spadj = PtrSize;
}
deltasp = autoffset;
if(debug['K'] > 1 && autoffset) {
// 6l -KK means double-check for stack overflow
// even after calling morestack and even if the
// function is marked as nosplit.
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_INDIR+D_CX;
p->from.offset = 0;
p->to.type = D_BX;
p = appendp(p);
p->as = ASUBQ;
p->from.type = D_CONST;
p->from.offset = StackSmall+32;
p->to.type = D_BX;
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_SP;
p->to.type = D_BX;
p = appendp(p);
p->as = AJHI;
p->to.type = D_BRANCH;
q1 = p;
p = appendp(p);
p->as = AINT;
p->from.type = D_CONST;
p->from.offset = 3;
p = appendp(p);
p->as = ANOP;
q1->pcond = p;
}
if(debug['Z'] && autoffset && !(cursym->text->from.scale&NOSPLIT)) {
// 6l -Z means zero the stack frame on entry.
// This slows down function calls but can help avoid
// false positives in garbage collection.
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_SP;
p->to.type = D_DI;
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_CONST;
p->from.offset = autoffset/8;
p->to.type = D_CX;
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_CONST;
p->from.offset = 0;
p->to.type = D_AX;
p = appendp(p);
p->as = AREP;
p = appendp(p);
p->as = ASTOSQ;
}
for(; p != P; p = p->link) {
pcsize = p->mode/8;
a = p->from.type;
if(a == D_AUTO)
p->from.offset += deltasp;
if(a == D_PARAM)
p->from.offset += deltasp + pcsize;
a = p->to.type;
if(a == D_AUTO)
p->to.offset += deltasp;
if(a == D_PARAM)
p->to.offset += deltasp + pcsize;
switch(p->as) {
default:
continue;
case APUSHL:
case APUSHFL:
deltasp += 4;
p->spadj = 4;
continue;
case APUSHQ:
case APUSHFQ:
deltasp += 8;
p->spadj = 8;
continue;
case APUSHW:
case APUSHFW:
deltasp += 2;
p->spadj = 2;
continue;
case APOPL:
case APOPFL:
deltasp -= 4;
p->spadj = -4;
continue;
case APOPQ:
case APOPFQ:
deltasp -= 8;
p->spadj = -8;
continue;
case APOPW:
case APOPFW:
deltasp -= 2;
p->spadj = -2;
continue;
case ARET:
break;
}
if(autoffset != deltasp)
diag("unbalanced PUSH/POP");
if(autoffset) {
p->as = AADJSP;
p->from.type = D_CONST;
p->from.offset = -autoffset;
p->spadj = -autoffset;
p = appendp(p);
p->as = ARET;
// If there are instructions following
// this ARET, they come from a branch
// with the same stackframe, so undo
// the cleanup.
p->spadj = +autoffset;
}
if(p->to.sym) // retjmp
p->as = AJMP;
}
}
}
void
dostkoff(void)
{
Prog *p, *q, *q1;
int32 autoffset, deltasp;
int a, pcsize;
uint32 moreconst1, moreconst2, i;
for(i=0; i<nelem(morename); i++) {
symmorestack[i] = lookup(morename[i], 0);
if(symmorestack[i]->type != STEXT)
diag("morestack trampoline not defined - %s", morename[i]);
pmorestack[i] = symmorestack[i]->text;
}
for(cursym = textp; cursym != nil; cursym = cursym->next) {
if(cursym->text == nil || cursym->text->link == nil)
continue;
p = cursym->text;
parsetextconst(p->to.offset);
autoffset = textstksiz;
if(autoffset < 0)
autoffset = 0;
if(autoffset < StackSmall && !(p->from.scale & NOSPLIT)) {
for(q = p; q != P; q = q->link)
if(q->as == ACALL)
goto noleaf;
p->from.scale |= NOSPLIT;
noleaf:;
}
q = P;
if((p->from.scale & NOSPLIT) && autoffset >= StackSmall)
diag("nosplit func likely to overflow stack");
if(!(p->from.scale & NOSPLIT)) {
p = appendp(p); // load g into CX
p->as = AMOVQ;
if(HEADTYPE == Hlinux || HEADTYPE == Hfreebsd
|| HEADTYPE == Hopenbsd || HEADTYPE == Hnetbsd
|| HEADTYPE == Hplan9x64) // ELF uses FS
p->from.type = D_INDIR+D_FS;
else
p->from.type = D_INDIR+D_GS;
p->from.offset = tlsoffset+0;
p->to.type = D_CX;
if(HEADTYPE == Hwindows) {
// movq %gs:0x28, %rcx
// movq (%rcx), %rcx
p->as = AMOVQ;
p->from.type = D_INDIR+D_GS;
p->from.offset = 0x28;
p->to.type = D_CX;
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_INDIR+D_CX;
p->from.offset = 0;
p->to.type = D_CX;
}
if(debug['K']) {
// 6l -K means check not only for stack
// overflow but stack underflow.
// On underflow, INT 3 (breakpoint).
// Underflow itself is rare but this also
// catches out-of-sync stack guard info
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_INDIR+D_CX;
p->from.offset = 8;
p->to.type = D_SP;
p = appendp(p);
p->as = AJHI;
p->to.type = D_BRANCH;
p->to.offset = 4;
q1 = p;
p = appendp(p);
p->as = AINT;
p->from.type = D_CONST;
p->from.offset = 3;
p = appendp(p);
p->as = ANOP;
q1->pcond = p;
}
if(autoffset < StackBig) { // do we need to call morestack?
if(autoffset <= StackSmall) {
// small stack
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_SP;
p->to.type = D_INDIR+D_CX;
} else {
// large stack
p = appendp(p);
p->as = ALEAQ;
p->from.type = D_INDIR+D_SP;
p->from.offset = -(autoffset-StackSmall);
p->to.type = D_AX;
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_AX;
p->to.type = D_INDIR+D_CX;
}
// common
p = appendp(p);
p->as = AJHI;
p->to.type = D_BRANCH;
p->to.offset = 4;
q = p;
}
// If we ask for more stack, we'll get a minimum of StackMin bytes.
// We need a stack frame large enough to hold the top-of-stack data,
// the function arguments+results, our caller's PC, our frame,
// a word for the return PC of the next call, and then the StackLimit bytes
// that must be available on entry to any function called from a function
// that did a stack check. If StackMin is enough, don't ask for a specific
// amount: then we can use the custom functions and save a few
// instructions.
moreconst1 = 0;
if(StackTop + textarg + PtrSize + autoffset + PtrSize + StackLimit >= StackMin)
moreconst1 = autoffset;
moreconst2 = textarg;
// 4 varieties varieties (const1==0 cross const2==0)
// and 6 subvarieties of (const1==0 and const2!=0)
p = appendp(p);
if(moreconst1 == 0 && moreconst2 == 0) {
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[0];
p->to.sym = symmorestack[0];
} else
if(moreconst1 != 0 && moreconst2 == 0) {
p->as = AMOVL;
p->from.type = D_CONST;
p->from.offset = moreconst1;
p->to.type = D_AX;
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[1];
p->to.sym = symmorestack[1];
} else
if(moreconst1 == 0 && moreconst2 <= 48 && moreconst2%8 == 0) {
i = moreconst2/8 + 3;
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[i];
p->to.sym = symmorestack[i];
} else
if(moreconst1 == 0 && moreconst2 != 0) {
p->as = AMOVL;
p->from.type = D_CONST;
p->from.offset = moreconst2;
p->to.type = D_AX;
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[2];
p->to.sym = symmorestack[2];
} else {
p->as = AMOVQ;
p->from.type = D_CONST;
p->from.offset = (uint64)moreconst2 << 32;
p->from.offset |= moreconst1;
p->to.type = D_AX;
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[3];
p->to.sym = symmorestack[3];
}
}
if(q != P)
q->pcond = p->link;
if(autoffset) {
p = appendp(p);
p->as = AADJSP;
p->from.type = D_CONST;
p->from.offset = autoffset;
p->spadj = autoffset;
if(q != P)
q->pcond = p;
} else {
// zero-byte stack adjustment.
// Insert a fake non-zero adjustment so that stkcheck can
// recognize the end of the stack-splitting prolog.
p = appendp(p);
p->as = ANOP;
p->spadj = -PtrSize;
p = appendp(p);
p->as = ANOP;
p->spadj = PtrSize;
}
deltasp = autoffset;
if(debug['K'] > 1 && autoffset) {
// 6l -KK means double-check for stack overflow
// even after calling morestack and even if the
// function is marked as nosplit.
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_INDIR+D_CX;
p->from.offset = 0;
p->to.type = D_BX;
p = appendp(p);
p->as = ASUBQ;
p->from.type = D_CONST;
p->from.offset = StackSmall+32;
p->to.type = D_BX;
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_SP;
p->to.type = D_BX;
p = appendp(p);
p->as = AJHI;
p->to.type = D_BRANCH;
q1 = p;
p = appendp(p);
p->as = AINT;
p->from.type = D_CONST;
p->from.offset = 3;
p = appendp(p);
p->as = ANOP;
q1->pcond = p;
}
if(debug['Z'] && autoffset && !(cursym->text->from.scale&NOSPLIT)) {
// 6l -Z means zero the stack frame on entry.
// This slows down function calls but can help avoid
// false positives in garbage collection.
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_SP;
p->to.type = D_DI;
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_CONST;
p->from.offset = autoffset/8;
p->to.type = D_CX;
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_CONST;
p->from.offset = 0;
p->to.type = D_AX;
p = appendp(p);
p->as = AREP;
p = appendp(p);
p->as = ASTOSQ;
}
for(; p != P; p = p->link) {
pcsize = p->mode/8;
a = p->from.type;
if(a == D_AUTO)
p->from.offset += deltasp;
if(a == D_PARAM)
p->from.offset += deltasp + pcsize;
a = p->to.type;
if(a == D_AUTO)
p->to.offset += deltasp;
if(a == D_PARAM)
p->to.offset += deltasp + pcsize;
switch(p->as) {
default:
continue;
case APUSHL:
case APUSHFL:
deltasp += 4;
p->spadj = 4;
continue;
case APUSHQ:
case APUSHFQ:
deltasp += 8;
p->spadj = 8;
continue;
case APUSHW:
case APUSHFW:
deltasp += 2;
p->spadj = 2;
continue;
case APOPL:
case APOPFL:
deltasp -= 4;
p->spadj = -4;
continue;
case APOPQ:
case APOPFQ:
deltasp -= 8;
p->spadj = -8;
continue;
case APOPW:
case APOPFW:
deltasp -= 2;
p->spadj = -2;
continue;
case ARET:
break;
}
if(autoffset != deltasp)
diag("unbalanced PUSH/POP");
if(autoffset) {
p->as = AADJSP;
p->from.type = D_CONST;
p->from.offset = -autoffset;
p->spadj = -autoffset;
p = appendp(p);
p->as = ARET;
// If there are instructions following
// this ARET, they come from a branch
// with the same stackframe, so undo
// the cleanup.
p->spadj = +autoffset;
}
}
}
}
void
dostkoff(void)
{
Prog *p, *q, *q1;
int32 autoffset, deltasp;
int a, f, curframe, curbecome, maxbecome, pcsize;
uint32 moreconst1, moreconst2, i;
for(i=0; i<nelem(morename); i++) {
symmorestack[i] = lookup(morename[i], 0);
pmorestack[i] = P;
}
for(p = firstp; p != P; p = p->link) {
if(p->as == ATEXT) {
for(i=0; i<nelem(morename); i++) {
if(p->from.sym == symmorestack[i]) {
pmorestack[i] = p;
break;
}
}
}
}
for(i=0; i<nelem(morename); i++) {
if(pmorestack[i] == P)
diag("morestack trampoline not defined");
}
curframe = 0;
curbecome = 0;
maxbecome = 0;
curtext = 0;
for(p = firstp; p != P; p = p->link) {
/* find out how much arg space is used in this TEXT */
if(p->to.type == (D_INDIR+D_SP))
if(p->to.offset > curframe)
curframe = p->to.offset;
switch(p->as) {
case ATEXT:
if(curtext && curtext->from.sym) {
curtext->from.sym->frame = curframe;
curtext->from.sym->become = curbecome;
if(curbecome > maxbecome)
maxbecome = curbecome;
}
curframe = 0;
curbecome = 0;
curtext = p;
break;
case ARET:
/* special form of RET is BECOME */
if(p->from.type == D_CONST)
if(p->from.offset > curbecome)
curbecome = p->from.offset;
break;
}
}
if(curtext && curtext->from.sym) {
curtext->from.sym->frame = curframe;
curtext->from.sym->become = curbecome;
if(curbecome > maxbecome)
maxbecome = curbecome;
}
if(debug['b'])
print("max become = %d\n", maxbecome);
xdefine("ALEFbecome", STEXT, maxbecome);
curtext = 0;
for(p = firstp; p != P; p = p->link) {
switch(p->as) {
case ATEXT:
curtext = p;
break;
case ACALL:
if(curtext != P && curtext->from.sym != S && curtext->to.offset >= 0) {
f = maxbecome - curtext->from.sym->frame;
if(f <= 0)
break;
/* calling a become or calling a variable */
if(p->to.sym == S || p->to.sym->become) {
curtext->to.offset += f;
if(debug['b']) {
curp = p;
print("%D calling %D increase %d\n",
&curtext->from, &p->to, f);
}
}
}
break;
}
}
autoffset = 0;
deltasp = 0;
for(p = firstp; p != P; p = p->link) {
if(p->as == ATEXT) {
curtext = p;
parsetextconst(p->to.offset);
autoffset = textstksiz;
if(autoffset < 0)
autoffset = 0;
q = P;
q1 = P;
if((p->from.scale & NOSPLIT) && autoffset >= StackSmall)
diag("nosplit func likely to overflow stack");
if(!(p->from.scale & NOSPLIT)) {
if(debug['K']) {
// 6l -K means check not only for stack
// overflow but stack underflow.
// On underflow, INT 3 (breakpoint).
// Underflow itself is rare but this also
// catches out-of-sync stack guard info
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_INDIR+D_R15;
p->from.offset = 8;
p->to.type = D_SP;
p = appendp(p);
p->as = AJHI;
p->to.type = D_BRANCH;
p->to.offset = 4;
q1 = p;
p = appendp(p);
p->as = AINT;
p->from.type = D_CONST;
p->from.offset = 3;
}
if(autoffset < StackBig) { // do we need to call morestack?
if(autoffset <= StackSmall) {
// small stack
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_SP;
p->to.type = D_INDIR+D_R15;
if(q1) {
q1->pcond = p;
q1 = P;
}
} else {
// large stack
p = appendp(p);
p->as = ALEAQ;
p->from.type = D_INDIR+D_SP;
p->from.offset = -(autoffset-StackSmall);
p->to.type = D_AX;
if(q1) {
q1->pcond = p;
q1 = P;
}
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_AX;
p->to.type = D_INDIR+D_R15;
}
// common
p = appendp(p);
p->as = AJHI;
p->to.type = D_BRANCH;
p->to.offset = 4;
q = p;
}
/* 160 comes from 3 calls (3*8) 4 safes (4*8) and 104 guard */
moreconst1 = 0;
if(autoffset+160 > 4096)
moreconst1 = (autoffset+160) & ~7LL;
moreconst2 = textarg;
// 4 varieties varieties (const1==0 cross const2==0)
// and 6 subvarieties of (const1==0 and const2!=0)
p = appendp(p);
if(moreconst1 == 0 && moreconst2 == 0) {
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[0];
p->to.sym = symmorestack[0];
if(q1) {
q1->pcond = p;
q1 = P;
}
} else
if(moreconst1 != 0 && moreconst2 == 0) {
p->as = AMOVL;
p->from.type = D_CONST;
p->from.offset = moreconst1;
p->to.type = D_AX;
if(q1) {
q1->pcond = p;
q1 = P;
}
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[1];
p->to.sym = symmorestack[1];
} else
if(moreconst1 == 0 && moreconst2 <= 48 && moreconst2%8 == 0) {
i = moreconst2/8 + 3;
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[i];
p->to.sym = symmorestack[i];
if(q1) {
q1->pcond = p;
q1 = P;
}
} else
if(moreconst1 == 0 && moreconst2 != 0) {
p->as = AMOVL;
p->from.type = D_CONST;
p->from.offset = moreconst2;
p->to.type = D_AX;
if(q1) {
q1->pcond = p;
q1 = P;
}
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[2];
p->to.sym = symmorestack[2];
} else {
p->as = AMOVQ;
p->from.type = D_CONST;
p->from.offset = (uint64)moreconst2 << 32;
p->from.offset |= moreconst1;
p->to.type = D_AX;
if(q1) {
q1->pcond = p;
q1 = P;
}
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
p->pcond = pmorestack[3];
p->to.sym = symmorestack[3];
}
}
if(q != P)
q->pcond = p->link;
if(autoffset) {
p = appendp(p);
p->as = AADJSP;
p->from.type = D_CONST;
p->from.offset = autoffset;
if(q != P)
q->pcond = p;
}
deltasp = autoffset;
if(debug['K'] > 1 && autoffset) {
// 6l -KK means double-check for stack overflow
// even after calling morestack and even if the
// function is marked as nosplit.
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_INDIR+D_R15;
p->from.offset = 0;
p->to.type = D_BX;
p = appendp(p);
p->as = ASUBQ;
p->from.type = D_CONST;
p->from.offset = StackSmall+32;
p->to.type = D_BX;
p = appendp(p);
p->as = ACMPQ;
p->from.type = D_SP;
p->to.type = D_BX;
p = appendp(p);
p->as = AJHI;
p->to.type = D_BRANCH;
q1 = p;
p = appendp(p);
p->as = AINT;
p->from.type = D_CONST;
p->from.offset = 3;
p = appendp(p);
p->as = ANOP;
q1->pcond = p;
q1 = P;
}
}
pcsize = p->mode/8;
a = p->from.type;
if(a == D_AUTO)
p->from.offset += deltasp;
if(a == D_PARAM)
p->from.offset += deltasp + pcsize;
a = p->to.type;
if(a == D_AUTO)
p->to.offset += deltasp;
if(a == D_PARAM)
p->to.offset += deltasp + pcsize;
switch(p->as) {
default:
continue;
case APUSHL:
case APUSHFL:
deltasp += 4;
continue;
case APUSHQ:
case APUSHFQ:
deltasp += 8;
continue;
case APUSHW:
case APUSHFW:
deltasp += 2;
continue;
case APOPL:
case APOPFL:
deltasp -= 4;
continue;
case APOPQ:
case APOPFQ:
deltasp -= 8;
continue;
case APOPW:
case APOPFW:
deltasp -= 2;
continue;
case ARET:
break;
}
if(autoffset != deltasp)
diag("unbalanced PUSH/POP");
if(p->from.type == D_CONST)
goto become;
if(autoffset) {
p->as = AADJSP;
p->from.type = D_CONST;
p->from.offset = -autoffset;
p = appendp(p);
p->as = ARET;
}
continue;
become:
q = p;
p = appendp(p);
p->as = AJMP;
p->to = q->to;
p->pcond = q->pcond;
q->as = AADJSP;
q->from = zprg.from;
q->from.type = D_CONST;
q->from.offset = -autoffset;
q->to = zprg.to;
continue;
}
}
