func stacksplit(ctxt *liblink.Link, p *liblink.Prog, framesize int32, noctxt int) *liblink.Prog { // MOVW g_stackguard(g), R1 p = liblink.Appendp(ctxt, p) p.As = AMOVW p.From.Type_ = D_OREG p.From.Reg = REGG p.From.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0 if ctxt.Cursym.Cfunc != 0 { p.From.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1 } p.To.Type_ = D_REG p.To.Reg = 1 if framesize <= liblink.StackSmall { // small stack: SP < stackguard // CMP stackguard, SP p = liblink.Appendp(ctxt, p) p.As = ACMP p.From.Type_ = D_REG p.From.Reg = 1 p.Reg = REGSP } else if framesize <= liblink.StackBig { // large stack: SP-framesize < stackguard-StackSmall // MOVW $-framesize(SP), R2 // CMP stackguard, R2 p = liblink.Appendp(ctxt, p) p.As = AMOVW p.From.Type_ = D_CONST p.From.Reg = REGSP p.From.Offset = int64(-framesize) p.To.Type_ = D_REG p.To.Reg = 2 p = liblink.Appendp(ctxt, p) p.As = ACMP p.From.Type_ = D_REG p.From.Reg = 1 p.Reg = 2 } 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. // CMP $StackPreempt, R1 // MOVW.NE $StackGuard(SP), R2 // SUB.NE R1, R2 // MOVW.NE $(framesize+(StackGuard-StackSmall)), R3 // CMP.NE R3, R2 p = liblink.Appendp(ctxt, p) p.As = ACMP p.From.Type_ = D_CONST p.From.Offset = int64(uint32(liblink.StackPreempt & (1<<32 - 1))) p.Reg = 1 p = liblink.Appendp(ctxt, p) p.As = AMOVW p.From.Type_ = D_CONST p.From.Reg = REGSP p.From.Offset = liblink.StackGuard p.To.Type_ = D_REG p.To.Reg = 2 p.Scond = C_SCOND_NE p = liblink.Appendp(ctxt, p) p.As = ASUB p.From.Type_ = D_REG p.From.Reg = 1 p.To.Type_ = D_REG p.To.Reg = 2 p.Scond = C_SCOND_NE p = liblink.Appendp(ctxt, p) p.As = AMOVW p.From.Type_ = D_CONST p.From.Offset = int64(framesize) + (liblink.StackGuard - liblink.StackSmall) p.To.Type_ = D_REG p.To.Reg = 3 p.Scond = C_SCOND_NE p = liblink.Appendp(ctxt, p) p.As = ACMP p.From.Type_ = D_REG p.From.Reg = 3 p.Reg = 2 p.Scond = C_SCOND_NE } // MOVW.LS R14, R3 p = liblink.Appendp(ctxt, p) p.As = AMOVW p.Scond = C_SCOND_LS p.From.Type_ = D_REG p.From.Reg = REGLINK p.To.Type_ = D_REG p.To.Reg = 3 // BL.LS runtime.morestack(SB) // modifies LR, returns with LO still asserted p = liblink.Appendp(ctxt, p) p.As = ABL p.Scond = C_SCOND_LS p.To.Type_ = D_BRANCH if ctxt.Cursym.Cfunc != 0 { p.To.Sym = liblink.Linklookup(ctxt, "runtime.morestackc", 0) } else { p.To.Sym = ctxt.Symmorestack[noctxt] } // BLS start p = liblink.Appendp(ctxt, p) p.As = ABLS p.To.Type_ = D_BRANCH p.Pcond = ctxt.Cursym.Text.Link return p }
func xfol(ctxt *liblink.Link, p *liblink.Prog, last **liblink.Prog) { var q *liblink.Prog var r *liblink.Prog var a int var i int loop: if p == nil { return } a = int(p.As) if a == AB { q = p.Pcond if q != nil && q.As != ATEXT { p.Mark |= FOLL p = q if !(p.Mark&FOLL != 0) { goto loop } } } if p.Mark&FOLL != 0 { i = 0 q = p for ; i < 4; (func() { i++; q = q.Link })() { if q == *last || q == nil { break } a = int(q.As) if a == ANOP { i-- continue } if a == AB || (a == ARET && q.Scond == C_SCOND_NONE) || a == ARFE || a == AUNDEF { goto copy } if q.Pcond == nil || (q.Pcond.Mark&FOLL != 0) { continue } if a != ABEQ && a != ABNE { continue } copy: for { r = ctxt.Arch.Prg() *r = *p if !(r.Mark&FOLL != 0) { fmt.Printf("can't happen 1\n") } r.Mark |= FOLL if p != q { p = p.Link (*last).Link = r *last = r continue } (*last).Link = r *last = r if a == AB || (a == ARET && q.Scond == C_SCOND_NONE) || a == ARFE || a == AUNDEF { return } r.As = ABNE if a == ABNE { r.As = ABEQ } r.Pcond = p.Link r.Link = p.Pcond if !(r.Link.Mark&FOLL != 0) { xfol(ctxt, r.Link, last) } if !(r.Pcond.Mark&FOLL != 0) { fmt.Printf("can't happen 2\n") } return } } a = AB q = ctxt.Arch.Prg() q.As = int16(a) q.Lineno = p.Lineno q.To.Type_ = D_BRANCH q.To.Offset = p.Pc q.Pcond = p p = q } p.Mark |= FOLL (*last).Link = p *last = p if a == AB || (a == ARET && p.Scond == C_SCOND_NONE) || a == ARFE || a == AUNDEF { return } if p.Pcond != nil { if a != ABL && a != ABX && p.Link != nil { q = liblink.Brchain(ctxt, p.Link) if a != ATEXT && a != ABCASE { if q != nil && (q.Mark&FOLL != 0) { p.As = int16(relinv(a)) p.Link = p.Pcond p.Pcond = q } } xfol(ctxt, p.Link, last) q = liblink.Brchain(ctxt, p.Pcond) if q == nil { q = p.Pcond } if q.Mark&FOLL != 0 { p.Pcond = q return } p = q goto loop } } p = p.Link goto loop }
func addstacksplit(ctxt *liblink.Link, cursym *liblink.LSym) { var p *liblink.Prog var pl *liblink.Prog var p1 *liblink.Prog var p2 *liblink.Prog var q *liblink.Prog var q1 *liblink.Prog var q2 *liblink.Prog var o int var autosize int32 var autoffset int32 autosize = 0 if ctxt.Symmorestack[0] == nil { ctxt.Symmorestack[0] = liblink.Linklookup(ctxt, "runtime.morestack", 0) ctxt.Symmorestack[1] = liblink.Linklookup(ctxt, "runtime.morestack_noctxt", 0) } q = nil ctxt.Cursym = cursym if cursym.Text == nil || cursym.Text.Link == nil { return } softfloat(ctxt, cursym) p = cursym.Text autoffset = int32(p.To.Offset) if autoffset < 0 { autoffset = 0 } cursym.Locals = autoffset cursym.Args = p.To.Offset2 if ctxt.Debugzerostack != 0 { if autoffset != 0 && !(p.Reg&liblink.NOSPLIT != 0) { // MOVW $4(R13), R1 p = liblink.Appendp(ctxt, p) p.As = AMOVW p.From.Type_ = D_CONST p.From.Reg = 13 p.From.Offset = 4 p.To.Type_ = D_REG p.To.Reg = 1 // MOVW $n(R13), R2 p = liblink.Appendp(ctxt, p) p.As = AMOVW p.From.Type_ = D_CONST p.From.Reg = 13 p.From.Offset = 4 + int64(autoffset) p.To.Type_ = D_REG p.To.Reg = 2 // MOVW $0, R3 p = liblink.Appendp(ctxt, p) p.As = AMOVW p.From.Type_ = D_CONST p.From.Offset = 0 p.To.Type_ = D_REG p.To.Reg = 3 // L: // MOVW.nil R3, 0(R1) +4 // CMP R1, R2 // BNE L pl = liblink.Appendp(ctxt, p) p = pl p.As = AMOVW p.From.Type_ = D_REG p.From.Reg = 3 p.To.Type_ = D_OREG p.To.Reg = 1 p.To.Offset = 4 p.Scond |= C_PBIT p = liblink.Appendp(ctxt, p) p.As = ACMP p.From.Type_ = D_REG p.From.Reg = 1 p.Reg = 2 p = liblink.Appendp(ctxt, p) p.As = ABNE p.To.Type_ = D_BRANCH p.Pcond = pl } } /* * find leaf subroutines * strip NOPs * expand RET * expand BECOME pseudo */ for p = cursym.Text; p != nil; p = p.Link { switch p.As { case ACASE: if ctxt.Flag_shared != 0 { linkcase(p) } case ATEXT: p.Mark |= LEAF case ARET: break case ADIV, ADIVU, AMOD, AMODU: q = p if ctxt.Sym_div == nil { initdiv(ctxt) } cursym.Text.Mark &^= LEAF continue case ANOP: q1 = p.Link q.Link = q1 /* q is non-nop */ if q1 != nil { q1.Mark |= p.Mark } continue case ABL, ABX, ADUFFZERO, ADUFFCOPY: cursym.Text.Mark &^= LEAF fallthrough case ABCASE, AB, ABEQ, ABNE, ABCS, ABHS, ABCC, ABLO, ABMI, ABPL, ABVS, ABVC, ABHI, ABLS, ABGE, ABLT, ABGT, ABLE: q1 = p.Pcond if q1 != nil { for q1.As == ANOP { q1 = q1.Link p.Pcond = q1 } } break } q = p } for p = cursym.Text; p != nil; p = p.Link { o = int(p.As) switch o { case ATEXT: autosize = int32(p.To.Offset + 4) if autosize <= 4 { if cursym.Text.Mark&LEAF != 0 { p.To.Offset = -4 autosize = 0 } } if !(autosize != 0) && !(cursym.Text.Mark&LEAF != 0) { if ctxt.Debugvlog != 0 { fmt.Fprintf(ctxt.Bso, "save suppressed in: %s\n", cursym.Name) liblink.Bflush(ctxt.Bso) } cursym.Text.Mark |= LEAF } if cursym.Text.Mark&LEAF != 0 { cursym.Leaf = 1 if !(autosize != 0) { break } } if !(p.Reg&liblink.NOSPLIT != 0) { p = stacksplit(ctxt, p, autosize, bool2int(!(cursym.Text.Reg&liblink.NEEDCTXT != 0))) // emit split check } // MOVW.W R14,$-autosize(SP) p = liblink.Appendp(ctxt, p) p.As = AMOVW p.Scond |= C_WBIT p.From.Type_ = D_REG p.From.Reg = REGLINK p.To.Type_ = D_OREG p.To.Offset = int64(-autosize) p.To.Reg = REGSP p.Spadj = autosize if cursym.Text.Reg&liblink.WRAPPER != 0 { // if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame // // MOVW g_panic(g), R1 // CMP $0, R1 // B.EQ end // MOVW panic_argp(R1), R2 // ADD $(autosize+4), R13, R3 // CMP R2, R3 // B.NE end // ADD $4, R13, R4 // MOVW R4, panic_argp(R1) // end: // NOP // // The NOP is needed to give the jumps somewhere to land. // It is a liblink NOP, not an ARM NOP: it encodes to 0 instruction bytes. p = liblink.Appendp(ctxt, p) p.As = AMOVW p.From.Type_ = D_OREG p.From.Reg = REGG p.From.Offset = 4 * int64(ctxt.Arch.Ptrsize) // G.panic p.To.Type_ = D_REG p.To.Reg = 1 p = liblink.Appendp(ctxt, p) p.As = ACMP p.From.Type_ = D_CONST p.From.Offset = 0 p.Reg = 1 p = liblink.Appendp(ctxt, p) p.As = ABEQ p.To.Type_ = D_BRANCH p1 = p p = liblink.Appendp(ctxt, p) p.As = AMOVW p.From.Type_ = D_OREG p.From.Reg = 1 p.From.Offset = 0 // Panic.argp p.To.Type_ = D_REG p.To.Reg = 2 p = liblink.Appendp(ctxt, p) p.As = AADD p.From.Type_ = D_CONST p.From.Offset = int64(autosize) + 4 p.Reg = 13 p.To.Type_ = D_REG p.To.Reg = 3 p = liblink.Appendp(ctxt, p) p.As = ACMP p.From.Type_ = D_REG p.From.Reg = 2 p.Reg = 3 p = liblink.Appendp(ctxt, p) p.As = ABNE p.To.Type_ = D_BRANCH p2 = p p = liblink.Appendp(ctxt, p) p.As = AADD p.From.Type_ = D_CONST p.From.Offset = 4 p.Reg = 13 p.To.Type_ = D_REG p.To.Reg = 4 p = liblink.Appendp(ctxt, p) p.As = AMOVW p.From.Type_ = D_REG p.From.Reg = 4 p.To.Type_ = D_OREG p.To.Reg = 1 p.To.Offset = 0 // Panic.argp p = liblink.Appendp(ctxt, p) p.As = ANOP p1.Pcond = p p2.Pcond = p } case ARET: nocache5(p) if cursym.Text.Mark&LEAF != 0 { if !(autosize != 0) { p.As = AB p.From = zprg5.From if p.To.Sym != nil { // retjmp p.To.Type_ = D_BRANCH } else { p.To.Type_ = D_OREG p.To.Offset = 0 p.To.Reg = REGLINK } break } } p.As = AMOVW p.Scond |= C_PBIT p.From.Type_ = D_OREG p.From.Offset = int64(autosize) p.From.Reg = REGSP p.To.Type_ = D_REG p.To.Reg = REGPC // If there are instructions following // this ARET, they come from a branch // with the same stackframe, so no spadj. if p.To.Sym != nil { // retjmp p.To.Reg = REGLINK q2 = liblink.Appendp(ctxt, p) q2.As = AB q2.To.Type_ = D_BRANCH q2.To.Sym = p.To.Sym p.To.Sym = nil p = q2 } case AADD: if p.From.Type_ == D_CONST && p.From.Reg == NREG && p.To.Type_ == D_REG && p.To.Reg == REGSP { p.Spadj = int32(-p.From.Offset) } case ASUB: if p.From.Type_ == D_CONST && p.From.Reg == NREG && p.To.Type_ == D_REG && p.To.Reg == REGSP { p.Spadj = int32(p.From.Offset) } case ADIV, ADIVU, AMOD, AMODU: if ctxt.Debugdivmod != 0 { break } if p.From.Type_ != D_REG { break } if p.To.Type_ != D_REG { break } q1 = p /* MOV a,4(SP) */ p = liblink.Appendp(ctxt, p) p.As = AMOVW p.Lineno = q1.Lineno p.From.Type_ = D_REG p.From.Reg = q1.From.Reg p.To.Type_ = D_OREG p.To.Reg = REGSP p.To.Offset = 4 /* MOV b,REGTMP */ p = liblink.Appendp(ctxt, p) p.As = AMOVW p.Lineno = q1.Lineno p.From.Type_ = D_REG p.From.Reg = int8(q1.Reg) if q1.Reg == NREG { p.From.Reg = q1.To.Reg } p.To.Type_ = D_REG p.To.Reg = REGTMP p.To.Offset = 0 /* CALL appropriate */ p = liblink.Appendp(ctxt, p) p.As = ABL p.Lineno = q1.Lineno p.To.Type_ = D_BRANCH switch o { case ADIV: p.To.Sym = ctxt.Sym_div case ADIVU: p.To.Sym = ctxt.Sym_divu case AMOD: p.To.Sym = ctxt.Sym_mod case AMODU: p.To.Sym = ctxt.Sym_modu break } /* MOV REGTMP, b */ p = liblink.Appendp(ctxt, p) p.As = AMOVW p.Lineno = q1.Lineno p.From.Type_ = D_REG p.From.Reg = REGTMP p.From.Offset = 0 p.To.Type_ = D_REG p.To.Reg = q1.To.Reg /* ADD $8,SP */ p = liblink.Appendp(ctxt, p) p.As = AADD p.Lineno = q1.Lineno p.From.Type_ = D_CONST p.From.Reg = NREG p.From.Offset = 8 p.Reg = NREG p.To.Type_ = D_REG p.To.Reg = REGSP p.Spadj = -8 /* Keep saved LR at 0(SP) after SP change. */ /* MOVW 0(SP), REGTMP; MOVW REGTMP, -8!(SP) */ /* TODO: Remove SP adjustments; see issue 6699. */ q1.As = AMOVW q1.From.Type_ = D_OREG q1.From.Reg = REGSP q1.From.Offset = 0 q1.Reg = NREG q1.To.Type_ = D_REG q1.To.Reg = REGTMP /* SUB $8,SP */ q1 = liblink.Appendp(ctxt, q1) q1.As = AMOVW q1.From.Type_ = D_REG q1.From.Reg = REGTMP q1.Reg = NREG q1.To.Type_ = D_OREG q1.To.Reg = REGSP q1.To.Offset = -8 q1.Scond |= C_WBIT q1.Spadj = 8 case AMOVW: if (p.Scond&C_WBIT != 0) && p.To.Type_ == D_OREG && p.To.Reg == REGSP { p.Spadj = int32(-p.To.Offset) } if (p.Scond&C_PBIT != 0) && p.From.Type_ == D_OREG && p.From.Reg == REGSP && p.To.Reg != REGPC { p.Spadj = int32(-p.From.Offset) } if p.From.Type_ == D_CONST && p.From.Reg == REGSP && p.To.Type_ == D_REG && p.To.Reg == REGSP { p.Spadj = int32(-p.From.Offset) } break } } }
/* // instruction scheduling if(debug['Q'] == 0) return; curtext = nil; q = nil; // p - 1 q1 = firstp; // top of block o = 0; // count of instructions for(p = firstp; p != nil; p = p1) { p1 = p->link; o++; if(p->mark & NOSCHED){ if(q1 != p){ sched(q1, q); } for(; p != nil; p = p->link){ if(!(p->mark & NOSCHED)) break; q = p; } p1 = p; q1 = p; o = 0; continue; } if(p->mark & (LABEL|SYNC)) { if(q1 != p) sched(q1, q); q1 = p; o = 1; } if(p->mark & (BRANCH|SYNC)) { sched(q1, p); q1 = p1; o = 0; } if(o >= NSCHED) { sched(q1, p); q1 = p1; o = 0; } q = p; } */ func stacksplit(ctxt *liblink.Link, p *liblink.Prog, framesize int32, noctxt int) *liblink.Prog { var q *liblink.Prog var q1 *liblink.Prog // MOVD g_stackguard(g), R3 p = liblink.Appendp(ctxt, p) p.As = AMOVD p.From.Type_ = D_OREG p.From.Reg = REGG p.From.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0 if ctxt.Cursym.Cfunc != 0 { p.From.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1 } p.To.Type_ = D_REG p.To.Reg = 3 q = nil if framesize <= liblink.StackSmall { // small stack: SP < stackguard // CMP stackguard, SP p = liblink.Appendp(ctxt, p) p.As = ACMPU p.From.Type_ = D_REG p.From.Reg = 3 p.To.Type_ = D_REG p.To.Reg = REGSP } else if framesize <= liblink.StackBig { // large stack: SP-framesize < stackguard-StackSmall // ADD $-framesize, SP, R4 // CMP stackguard, R4 p = liblink.Appendp(ctxt, p) p.As = AADD p.From.Type_ = D_CONST p.From.Offset = int64(-framesize) p.Reg = REGSP p.To.Type_ = D_REG p.To.Reg = 4 p = liblink.Appendp(ctxt, p) p.As = ACMPU p.From.Type_ = D_REG p.From.Reg = 3 p.To.Type_ = D_REG p.To.Reg = 4 } 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. // // stackguard is R3 // CMP R3, $StackPreempt // BEQ label-of-call-to-morestack // ADD $StackGuard, SP, R4 // SUB R3, R4 // MOVD $(framesize+(StackGuard-StackSmall)), R31 // CMPU R31, R4 p = liblink.Appendp(ctxt, p) p.As = ACMP p.From.Type_ = D_REG p.From.Reg = 3 p.To.Type_ = D_CONST p.To.Offset = liblink.StackPreempt p = liblink.Appendp(ctxt, p) q = p p.As = ABEQ p.To.Type_ = D_BRANCH p = liblink.Appendp(ctxt, p) p.As = AADD p.From.Type_ = D_CONST p.From.Offset = liblink.StackGuard p.Reg = REGSP p.To.Type_ = D_REG p.To.Reg = 4 p = liblink.Appendp(ctxt, p) p.As = ASUB p.From.Type_ = D_REG p.From.Reg = 3 p.To.Type_ = D_REG p.To.Reg = 4 p = liblink.Appendp(ctxt, p) p.As = AMOVD p.From.Type_ = D_CONST p.From.Offset = int64(framesize) + liblink.StackGuard - liblink.StackSmall p.To.Type_ = D_REG p.To.Reg = REGTMP p = liblink.Appendp(ctxt, p) p.As = ACMPU p.From.Type_ = D_REG p.From.Reg = REGTMP p.To.Type_ = D_REG p.To.Reg = 4 } // q1: BLT done p = liblink.Appendp(ctxt, p) q1 = p p.As = ABLT p.To.Type_ = D_BRANCH // MOVD LR, R5 p = liblink.Appendp(ctxt, p) p.As = AMOVD p.From.Type_ = D_SPR p.From.Offset = D_LR p.To.Type_ = D_REG p.To.Reg = 5 if q != nil { q.Pcond = p } // BL runtime.morestack(SB) p = liblink.Appendp(ctxt, p) p.As = ABL p.To.Type_ = D_BRANCH if ctxt.Cursym.Cfunc != 0 { p.To.Sym = liblink.Linklookup(ctxt, "runtime.morestackc", 0) } else { p.To.Sym = ctxt.Symmorestack[noctxt] } // BR start p = liblink.Appendp(ctxt, p) p.As = ABR p.To.Type_ = D_BRANCH p.Pcond = ctxt.Cursym.Text.Link // placeholder for q1's jump target p = liblink.Appendp(ctxt, p) p.As = ANOP // zero-width place holder q1.Pcond = p return p }
func xfol(ctxt *liblink.Link, p *liblink.Prog, last **liblink.Prog) { var q *liblink.Prog var r *liblink.Prog var a int var b int var i int loop: if p == nil { return } a = int(p.As) if a == ABR { q = p.Pcond if (p.Mark&NOSCHED != 0) || q != nil && (q.Mark&NOSCHED != 0) { p.Mark |= FOLL (*last).Link = p *last = p p = p.Link xfol(ctxt, p, last) p = q if p != nil && !(p.Mark&FOLL != 0) { goto loop } return } if q != nil { p.Mark |= FOLL p = q if !(p.Mark&FOLL != 0) { goto loop } } } if p.Mark&FOLL != 0 { i = 0 q = p for ; i < 4; (func() { i++; q = q.Link })() { if q == *last || (q.Mark&NOSCHED != 0) { break } b = 0 /* set */ a = int(q.As) if a == ANOP { i-- continue } if a == ABR || a == ARETURN || a == ARFI || a == ARFCI || a == ARFID || a == AHRFID { goto copy } if !(q.Pcond != nil) || (q.Pcond.Mark&FOLL != 0) { continue } b = relinv(a) if !(b != 0) { continue } copy: for { r = ctxt.Arch.Prg() *r = *p if !(r.Mark&FOLL != 0) { fmt.Printf("cant happen 1\n") } r.Mark |= FOLL if p != q { p = p.Link (*last).Link = r *last = r continue } (*last).Link = r *last = r if a == ABR || a == ARETURN || a == ARFI || a == ARFCI || a == ARFID || a == AHRFID { return } r.As = int16(b) r.Pcond = p.Link r.Link = p.Pcond if !(r.Link.Mark&FOLL != 0) { xfol(ctxt, r.Link, last) } if !(r.Pcond.Mark&FOLL != 0) { fmt.Printf("cant happen 2\n") } return } } a = ABR q = ctxt.Arch.Prg() q.As = int16(a) q.Lineno = p.Lineno q.To.Type_ = D_BRANCH q.To.Offset = p.Pc q.Pcond = p p = q } p.Mark |= FOLL (*last).Link = p *last = p if a == ABR || a == ARETURN || a == ARFI || a == ARFCI || a == ARFID || a == AHRFID { if p.Mark&NOSCHED != 0 { p = p.Link goto loop } return } if p.Pcond != nil { if a != ABL && p.Link != nil { xfol(ctxt, p.Link, last) p = p.Pcond if p == nil || (p.Mark&FOLL != 0) { return } goto loop } } p = p.Link goto loop }
func addstacksplit(ctxt *liblink.Link, cursym *liblink.LSym) { var p *liblink.Prog var q *liblink.Prog var p1 *liblink.Prog var p2 *liblink.Prog var q1 *liblink.Prog var o int var mov int var aoffset int var textstksiz int64 var textarg int64 var autosize int32 if ctxt.Symmorestack[0] == nil { ctxt.Symmorestack[0] = liblink.Linklookup(ctxt, "runtime.morestack", 0) ctxt.Symmorestack[1] = liblink.Linklookup(ctxt, "runtime.morestack_noctxt", 0) } // TODO(minux): add morestack short-cuts with small fixed frame-size. ctxt.Cursym = cursym if cursym.Text == nil || cursym.Text.Link == nil { return } p = cursym.Text parsetextconst(p.To.Offset, &textstksiz, &textarg) cursym.Args = int32(p.To.Offset >> 32) cursym.Locals = int32(textstksiz) /* * find leaf subroutines * strip NOPs * expand RET * expand BECOME pseudo */ if ctxt.Debugvlog != 0 { fmt.Fprintf(ctxt.Bso, "%5.2f noops\n", liblink.Cputime()) } liblink.Bflush(ctxt.Bso) q = nil for p = cursym.Text; p != nil; p = p.Link { switch p.As { /* too hard, just leave alone */ case ATEXT: q = p p.Mark |= LABEL | LEAF | SYNC if p.Link != nil { p.Link.Mark |= LABEL } case ANOR: q = p if p.To.Type_ == D_REG { if p.To.Reg == REGZERO { p.Mark |= LABEL | SYNC } } case ALWAR, ASTWCCC, AECIWX, AECOWX, AEIEIO, AICBI, AISYNC, ATLBIE, ATLBIEL, ASLBIA, ASLBIE, ASLBMFEE, ASLBMFEV, ASLBMTE, ADCBF, ADCBI, ADCBST, ADCBT, ADCBTST, ADCBZ, ASYNC, ATLBSYNC, APTESYNC, ATW, AWORD, ARFI, ARFCI, ARFID, AHRFID: q = p p.Mark |= LABEL | SYNC continue case AMOVW, AMOVWZ, AMOVD: q = p switch p.From.Type_ { case D_MSR, D_SPR, D_FPSCR, D_CREG, D_DCR: p.Mark |= LABEL | SYNC } switch p.To.Type_ { case D_MSR, D_SPR, D_FPSCR, D_CREG, D_DCR: p.Mark |= LABEL | SYNC } continue case AFABS, AFABSCC, AFADD, AFADDCC, AFCTIW, AFCTIWCC, AFCTIWZ, AFCTIWZCC, AFDIV, AFDIVCC, AFMADD, AFMADDCC, AFMOVD, AFMOVDU, /* case AFMOVDS: */ AFMOVS, AFMOVSU, /* case AFMOVSD: */ AFMSUB, AFMSUBCC, AFMUL, AFMULCC, AFNABS, AFNABSCC, AFNEG, AFNEGCC, AFNMADD, AFNMADDCC, AFNMSUB, AFNMSUBCC, AFRSP, AFRSPCC, AFSUB, AFSUBCC: q = p p.Mark |= FLOAT continue case ABL, ABCL, ADUFFZERO, ADUFFCOPY: cursym.Text.Mark &^= LEAF fallthrough case ABC, ABEQ, ABGE, ABGT, ABLE, ABLT, ABNE, ABR, ABVC, ABVS: p.Mark |= BRANCH q = p q1 = p.Pcond if q1 != nil { for q1.As == ANOP { q1 = q1.Link p.Pcond = q1 } if !(q1.Mark&LEAF != 0) { q1.Mark |= LABEL } } else { p.Mark |= LABEL } q1 = p.Link if q1 != nil { q1.Mark |= LABEL } continue case AFCMPO, AFCMPU: q = p p.Mark |= FCMP | FLOAT continue case ARETURN: q = p if p.Link != nil { p.Link.Mark |= LABEL } continue case ANOP: q1 = p.Link q.Link = q1 /* q is non-nop */ q1.Mark |= p.Mark continue default: q = p continue } } autosize = 0 for p = cursym.Text; p != nil; p = p.Link { o = int(p.As) switch o { case ATEXT: mov = AMOVD aoffset = 0 autosize = int32(textstksiz + 8) if (p.Mark&LEAF != 0) && autosize <= 8 { autosize = 0 } else if autosize&4 != 0 { autosize += 4 } p.To.Offset = int64(uint64(p.To.Offset)&(0xffffffff<<32) | uint64(uint32(autosize-8))) if !(p.Reg&liblink.NOSPLIT != 0) { p = stacksplit(ctxt, p, autosize, bool2int(!(cursym.Text.Reg&liblink.NEEDCTXT != 0))) // emit split check } q = p if autosize != 0 { /* use MOVDU to adjust R1 when saving R31, if autosize is small */ if !(cursym.Text.Mark&LEAF != 0) && autosize >= -BIG && autosize <= BIG { mov = AMOVDU aoffset = int(-autosize) } else { q = liblink.Appendp(ctxt, p) q.As = AADD q.Lineno = p.Lineno q.From.Type_ = D_CONST q.From.Offset = int64(-autosize) q.To.Type_ = D_REG q.To.Reg = REGSP q.Spadj = +autosize } } else if !(cursym.Text.Mark&LEAF != 0) { if ctxt.Debugvlog != 0 { fmt.Fprintf(ctxt.Bso, "save suppressed in: %s\n", cursym.Name) liblink.Bflush(ctxt.Bso) } cursym.Text.Mark |= LEAF } if cursym.Text.Mark&LEAF != 0 { cursym.Leaf = 1 break } q = liblink.Appendp(ctxt, q) q.As = AMOVD q.Lineno = p.Lineno q.From.Type_ = D_SPR q.From.Offset = D_LR q.To.Type_ = D_REG q.To.Reg = REGTMP q = liblink.Appendp(ctxt, q) q.As = int16(mov) q.Lineno = p.Lineno q.From.Type_ = D_REG q.From.Reg = REGTMP q.To.Type_ = D_OREG q.To.Offset = int64(aoffset) q.To.Reg = REGSP if q.As == AMOVDU { q.Spadj = int32(-aoffset) } if cursym.Text.Reg&liblink.WRAPPER != 0 { // if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame // // MOVD g_panic(g), R3 // CMP R0, R3 // BEQ end // MOVD panic_argp(R3), R4 // ADD $(autosize+8), R1, R5 // CMP R4, R5 // BNE end // ADD $8, R1, R6 // MOVD R6, panic_argp(R3) // end: // NOP // // The NOP is needed to give the jumps somewhere to land. // It is a liblink NOP, not a ppc64 NOP: it encodes to 0 instruction bytes. q = liblink.Appendp(ctxt, q) q.As = AMOVD q.From.Type_ = D_OREG q.From.Reg = REGG q.From.Offset = 4 * int64(ctxt.Arch.Ptrsize) // G.panic q.To.Type_ = D_REG q.To.Reg = 3 q = liblink.Appendp(ctxt, q) q.As = ACMP q.From.Type_ = D_REG q.From.Reg = 0 q.To.Type_ = D_REG q.To.Reg = 3 q = liblink.Appendp(ctxt, q) q.As = ABEQ q.To.Type_ = D_BRANCH p1 = q q = liblink.Appendp(ctxt, q) q.As = AMOVD q.From.Type_ = D_OREG q.From.Reg = 3 q.From.Offset = 0 // Panic.argp q.To.Type_ = D_REG q.To.Reg = 4 q = liblink.Appendp(ctxt, q) q.As = AADD q.From.Type_ = D_CONST q.From.Offset = int64(autosize) + 8 q.Reg = REGSP q.To.Type_ = D_REG q.To.Reg = 5 q = liblink.Appendp(ctxt, q) q.As = ACMP q.From.Type_ = D_REG q.From.Reg = 4 q.To.Type_ = D_REG q.To.Reg = 5 q = liblink.Appendp(ctxt, q) q.As = ABNE q.To.Type_ = D_BRANCH p2 = q q = liblink.Appendp(ctxt, q) q.As = AADD q.From.Type_ = D_CONST q.From.Offset = 8 q.Reg = REGSP q.To.Type_ = D_REG q.To.Reg = 6 q = liblink.Appendp(ctxt, q) q.As = AMOVD q.From.Type_ = D_REG q.From.Reg = 6 q.To.Type_ = D_OREG q.To.Reg = 3 q.To.Offset = 0 // Panic.argp q = liblink.Appendp(ctxt, q) q.As = ANOP p1.Pcond = q p2.Pcond = q } case ARETURN: if p.From.Type_ == D_CONST { ctxt.Diag("using BECOME (%P) is not supported!", p) break } if p.To.Sym != nil { // retjmp p.As = ABR p.To.Type_ = D_BRANCH break } if cursym.Text.Mark&LEAF != 0 { if !(autosize != 0) { p.As = ABR p.From = zprg.From p.To.Type_ = D_SPR p.To.Offset = D_LR p.Mark |= BRANCH break } p.As = AADD p.From.Type_ = D_CONST p.From.Offset = int64(autosize) p.To.Type_ = D_REG p.To.Reg = REGSP p.Spadj = -autosize q = ctxt.Arch.Prg() q.As = ABR q.Lineno = p.Lineno q.To.Type_ = D_SPR q.To.Offset = D_LR q.Mark |= BRANCH q.Spadj = +autosize q.Link = p.Link p.Link = q break } p.As = AMOVD p.From.Type_ = D_OREG p.From.Offset = 0 p.From.Reg = REGSP p.To.Type_ = D_REG p.To.Reg = REGTMP q = ctxt.Arch.Prg() q.As = AMOVD q.Lineno = p.Lineno q.From.Type_ = D_REG q.From.Reg = REGTMP q.To.Type_ = D_SPR q.To.Offset = D_LR q.Link = p.Link p.Link = q p = q if false { // Debug bad returns q = ctxt.Arch.Prg() q.As = AMOVD q.Lineno = p.Lineno q.From.Type_ = D_OREG q.From.Offset = 0 q.From.Reg = REGTMP q.To.Type_ = D_REG q.To.Reg = REGTMP q.Link = p.Link p.Link = q p = q } if autosize != 0 { q = ctxt.Arch.Prg() q.As = AADD q.Lineno = p.Lineno q.From.Type_ = D_CONST q.From.Offset = int64(autosize) q.To.Type_ = D_REG q.To.Reg = REGSP q.Spadj = -autosize q.Link = p.Link p.Link = q } q1 = ctxt.Arch.Prg() q1.As = ABR q1.Lineno = p.Lineno q1.To.Type_ = D_SPR q1.To.Offset = D_LR q1.Mark |= BRANCH q1.Spadj = +autosize q1.Link = q.Link q.Link = q1 case AADD: if p.To.Type_ == D_REG && p.To.Reg == REGSP && p.From.Type_ == D_CONST { p.Spadj = int32(-p.From.Offset) } break } } }
func xfol(ctxt *liblink.Link, p *liblink.Prog, last **liblink.Prog) { var q *liblink.Prog var i int var a int loop: if p == nil { return } if p.As == AJMP { q = p.Pcond if q != nil && q.As != ATEXT { /* mark instruction as done and continue layout at target of jump */ p.Mark = 1 p = q if p.Mark == 0 { goto loop } } } if p.Mark != 0 { /* * p goes here, but already used it elsewhere. * copy up to 4 instructions or else branch to other copy. */ i = 0 q = p for ; i < 4; (func() { i++; q = q.Link })() { if q == nil { break } if q == *last { break } a = int(q.As) if a == ANOP { i-- continue } if nofollow(a) != 0 || pushpop(a) != 0 { break // NOTE(rsc): arm does goto copy } if q.Pcond == nil || q.Pcond.Mark != 0 { continue } if a == ACALL || a == ALOOP { continue } for { if p.As == ANOP { p = p.Link continue } q = liblink.Copyp(ctxt, p) p = p.Link q.Mark = 1 (*last).Link = q *last = q if int(q.As) != a || q.Pcond == nil || q.Pcond.Mark != 0 { continue } q.As = int16(relinv(int(q.As))) p = q.Pcond q.Pcond = q.Link q.Link = p xfol(ctxt, q.Link, last) p = q.Link if p.Mark != 0 { return } goto loop /* */ } } q = ctxt.Arch.Prg() q.As = AJMP q.Lineno = p.Lineno q.To.Type_ = D_BRANCH q.To.Offset = p.Pc q.Pcond = p p = q } /* emit p */ p.Mark = 1 (*last).Link = p *last = p a = int(p.As) /* continue loop with what comes after p */ if nofollow(a) != 0 { return } if p.Pcond != nil && a != ACALL { /* * some kind of conditional branch. * recurse to follow one path. * continue loop on the other. */ q = liblink.Brchain(ctxt, p.Pcond) if q != nil { p.Pcond = q } q = liblink.Brchain(ctxt, p.Link) if q != nil { p.Link = q } if p.From.Type_ == D_CONST { if p.From.Offset == 1 { /* * expect conditional jump to be taken. * rewrite so that's the fall-through case. */ p.As = int16(relinv(a)) q = p.Link p.Link = p.Pcond p.Pcond = q } } else { q = p.Link if q.Mark != 0 { if a != ALOOP { p.As = int16(relinv(a)) p.Link = p.Pcond p.Pcond = q } } } xfol(ctxt, p.Link, last) if p.Pcond.Mark != 0 { return } p = p.Pcond goto loop } p = p.Link goto loop }
// Append code to p to check for stack split. // Appends to (does not overwrite) p. // Assumes g is in CX. // Returns last new instruction. // On return, *jmpok is the instruction that should jump // to the stack frame allocation if no split is needed. func stacksplit(ctxt *liblink.Link, p *liblink.Prog, framesize int32, noctxt int, jmpok **liblink.Prog) *liblink.Prog { var q *liblink.Prog var q1 *liblink.Prog if ctxt.Debugstack != 0 { // 8l -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 = liblink.Appendp(ctxt, p) p.As = ACMPL p.From.Type_ = D_INDIR + D_CX p.From.Offset = 4 p.To.Type_ = D_SP p = liblink.Appendp(ctxt, p) p.As = AJCC p.To.Type_ = D_BRANCH p.To.Offset = 4 q1 = p p = liblink.Appendp(ctxt, p) p.As = AINT p.From.Type_ = D_CONST p.From.Offset = 3 p = liblink.Appendp(ctxt, p) p.As = ANOP q1.Pcond = p } q1 = nil if framesize <= liblink.StackSmall { // small stack: SP <= stackguard // CMPL SP, stackguard p = liblink.Appendp(ctxt, p) p.As = ACMPL p.From.Type_ = D_SP p.To.Type_ = D_INDIR + D_CX p.To.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0 if ctxt.Cursym.Cfunc != 0 { p.To.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1 } } else if framesize <= liblink.StackBig { // large stack: SP-framesize <= stackguard-StackSmall // LEAL -(framesize-StackSmall)(SP), AX // CMPL AX, stackguard p = liblink.Appendp(ctxt, p) p.As = ALEAL p.From.Type_ = D_INDIR + D_SP p.From.Offset = -(int64(framesize) - liblink.StackSmall) p.To.Type_ = D_AX p = liblink.Appendp(ctxt, p) p.As = ACMPL p.From.Type_ = D_AX p.To.Type_ = D_INDIR + D_CX p.To.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0 if ctxt.Cursym.Cfunc != 0 { p.To.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1 } } 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. // MOVL stackguard, CX // CMPL CX, $StackPreempt // JEQ label-of-call-to-morestack // LEAL StackGuard(SP), AX // SUBL stackguard, AX // CMPL AX, $(framesize+(StackGuard-StackSmall)) p = liblink.Appendp(ctxt, p) p.As = AMOVL p.From.Type_ = D_INDIR + D_CX p.From.Offset = 0 p.From.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0 if ctxt.Cursym.Cfunc != 0 { p.From.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1 } p.To.Type_ = D_SI p = liblink.Appendp(ctxt, p) p.As = ACMPL p.From.Type_ = D_SI p.To.Type_ = D_CONST p.To.Offset = int64(uint32(liblink.StackPreempt & (1<<32 - 1))) p = liblink.Appendp(ctxt, p) p.As = AJEQ p.To.Type_ = D_BRANCH q1 = p p = liblink.Appendp(ctxt, p) p.As = ALEAL p.From.Type_ = D_INDIR + D_SP p.From.Offset = liblink.StackGuard p.To.Type_ = D_AX p = liblink.Appendp(ctxt, p) p.As = ASUBL p.From.Type_ = D_SI p.From.Offset = 0 p.To.Type_ = D_AX p = liblink.Appendp(ctxt, p) p.As = ACMPL p.From.Type_ = D_AX p.To.Type_ = D_CONST p.To.Offset = int64(framesize) + (liblink.StackGuard - liblink.StackSmall) } // common p = liblink.Appendp(ctxt, p) p.As = AJHI p.To.Type_ = D_BRANCH p.To.Offset = 4 q = p p = liblink.Appendp(ctxt, p) p.As = ACALL p.To.Type_ = D_BRANCH if ctxt.Cursym.Cfunc != 0 { p.To.Sym = liblink.Linklookup(ctxt, "runtime.morestackc", 0) } else { p.To.Sym = ctxt.Symmorestack[noctxt] } p = liblink.Appendp(ctxt, p) p.As = AJMP p.To.Type_ = D_BRANCH p.Pcond = ctxt.Cursym.Text.Link if q != nil { q.Pcond = p.Link } if q1 != nil { q1.Pcond = q.Link } *jmpok = q return p }
func addstacksplit(ctxt *liblink.Link, cursym *liblink.LSym) { var p *liblink.Prog var q *liblink.Prog var p1 *liblink.Prog var p2 *liblink.Prog var autoffset int32 var deltasp int32 var a int if ctxt.Symmorestack[0] == nil { ctxt.Symmorestack[0] = liblink.Linklookup(ctxt, "runtime.morestack", 0) ctxt.Symmorestack[1] = liblink.Linklookup(ctxt, "runtime.morestack_noctxt", 0) } if ctxt.Headtype == liblink.Hplan9 && ctxt.Plan9privates == nil { ctxt.Plan9privates = liblink.Linklookup(ctxt, "_privates", 0) } ctxt.Cursym = cursym if cursym.Text == nil || cursym.Text.Link == nil { return } p = cursym.Text autoffset = int32(p.To.Offset) if autoffset < 0 { autoffset = 0 } cursym.Locals = autoffset cursym.Args = p.To.Offset2 q = nil if !(p.From.Scale&liblink.NOSPLIT != 0) || (p.From.Scale&liblink.WRAPPER != 0) { p = liblink.Appendp(ctxt, p) p = load_g_cx(ctxt, p) // load g into CX } if !(cursym.Text.From.Scale&liblink.NOSPLIT != 0) { p = stacksplit(ctxt, p, autoffset, bool2int(!(cursym.Text.From.Scale&liblink.NEEDCTXT != 0)), &q) // emit split check } if autoffset != 0 { p = liblink.Appendp(ctxt, p) p.As = AADJSP p.From.Type_ = D_CONST p.From.Offset = int64(autoffset) p.Spadj = autoffset } else { // zero-byte stack adjustment. // Insert a fake non-zero adjustment so that stkcheck can // recognize the end of the stack-splitting prolog. p = liblink.Appendp(ctxt, p) p.As = ANOP p.Spadj = int32(-ctxt.Arch.Ptrsize) p = liblink.Appendp(ctxt, p) p.As = ANOP p.Spadj = int32(ctxt.Arch.Ptrsize) } if q != nil { q.Pcond = p } deltasp = autoffset if cursym.Text.From.Scale&liblink.WRAPPER != 0 { // if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame // // MOVL g_panic(CX), BX // TESTL BX, BX // JEQ end // LEAL (autoffset+4)(SP), DI // CMPL panic_argp(BX), DI // JNE end // MOVL SP, panic_argp(BX) // end: // NOP // // The NOP is needed to give the jumps somewhere to land. // It is a liblink NOP, not an x86 NOP: it encodes to 0 instruction bytes. p = liblink.Appendp(ctxt, p) p.As = AMOVL p.From.Type_ = D_INDIR + D_CX p.From.Offset = 4 * int64(ctxt.Arch.Ptrsize) // G.panic p.To.Type_ = D_BX p = liblink.Appendp(ctxt, p) p.As = ATESTL p.From.Type_ = D_BX p.To.Type_ = D_BX p = liblink.Appendp(ctxt, p) p.As = AJEQ p.To.Type_ = D_BRANCH p1 = p p = liblink.Appendp(ctxt, p) p.As = ALEAL p.From.Type_ = D_INDIR + D_SP p.From.Offset = int64(autoffset) + 4 p.To.Type_ = D_DI p = liblink.Appendp(ctxt, p) p.As = ACMPL p.From.Type_ = D_INDIR + D_BX p.From.Offset = 0 // Panic.argp p.To.Type_ = D_DI p = liblink.Appendp(ctxt, p) p.As = AJNE p.To.Type_ = D_BRANCH p2 = p p = liblink.Appendp(ctxt, p) p.As = AMOVL p.From.Type_ = D_SP p.To.Type_ = D_INDIR + D_BX p.To.Offset = 0 // Panic.argp p = liblink.Appendp(ctxt, p) p.As = ANOP p1.Pcond = p p2.Pcond = p } if ctxt.Debugzerostack != 0 && autoffset != 0 && !(cursym.Text.From.Scale&liblink.NOSPLIT != 0) { // 8l -Z means zero the stack frame on entry. // This slows down function calls but can help avoid // false positives in garbage collection. p = liblink.Appendp(ctxt, p) p.As = AMOVL p.From.Type_ = D_SP p.To.Type_ = D_DI p = liblink.Appendp(ctxt, p) p.As = AMOVL p.From.Type_ = D_CONST p.From.Offset = int64(autoffset) / 4 p.To.Type_ = D_CX p = liblink.Appendp(ctxt, p) p.As = AMOVL p.From.Type_ = D_CONST p.From.Offset = 0 p.To.Type_ = D_AX p = liblink.Appendp(ctxt, p) p.As = AREP p = liblink.Appendp(ctxt, p) p.As = ASTOSL } for ; p != nil; p = p.Link { a = int(p.From.Type_) if a == D_AUTO { p.From.Offset += int64(deltasp) } if a == D_PARAM { p.From.Offset += int64(deltasp) + 4 } a = int(p.To.Type_) if a == D_AUTO { p.To.Offset += int64(deltasp) } if a == D_PARAM { p.To.Offset += int64(deltasp) + 4 } switch p.As { default: continue case APUSHL, APUSHFL: deltasp += 4 p.Spadj = 4 continue case APUSHW, APUSHFW: deltasp += 2 p.Spadj = 2 continue case APOPL, APOPFL: deltasp -= 4 p.Spadj = -4 continue case APOPW, APOPFW: deltasp -= 2 p.Spadj = -2 continue case ARET: break } if autoffset != deltasp { ctxt.Diag("unbalanced PUSH/POP") } if autoffset != 0 { p.As = AADJSP p.From.Type_ = D_CONST p.From.Offset = int64(-autoffset) p.Spadj = -autoffset p = liblink.Appendp(ctxt, 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 != nil { // retjmp p.As = AJMP } } }
func addstacksplit(ctxt *liblink.Link, cursym *liblink.LSym) { var p *liblink.Prog var q *liblink.Prog var p1 *liblink.Prog var p2 *liblink.Prog var autoffset int32 var deltasp int32 var a int var pcsize int var textstksiz int64 var textarg int64 if ctxt.Tlsg == nil { ctxt.Tlsg = liblink.Linklookup(ctxt, "runtime.tlsg", 0) } if ctxt.Symmorestack[0] == nil { ctxt.Symmorestack[0] = liblink.Linklookup(ctxt, "runtime.morestack", 0) ctxt.Symmorestack[1] = liblink.Linklookup(ctxt, "runtime.morestack_noctxt", 0) } if ctxt.Headtype == liblink.Hplan9 && ctxt.Plan9privates == nil { ctxt.Plan9privates = liblink.Linklookup(ctxt, "_privates", 0) } ctxt.Cursym = cursym if cursym.Text == nil || cursym.Text.Link == nil { return } p = cursym.Text parsetextconst(p.To.Offset, &textstksiz, &textarg) autoffset = int32(textstksiz) if autoffset < 0 { autoffset = 0 } cursym.Args = int32(p.To.Offset >> 32) cursym.Locals = int32(textstksiz) if autoffset < liblink.StackSmall && !(p.From.Scale&liblink.NOSPLIT != 0) { for q = p; q != nil; q = q.Link { if q.As == ACALL { goto noleaf } if (q.As == ADUFFCOPY || q.As == ADUFFZERO) && autoffset >= liblink.StackSmall-8 { goto noleaf } } p.From.Scale |= liblink.NOSPLIT noleaf: } q = nil if !(p.From.Scale&liblink.NOSPLIT != 0) || (p.From.Scale&liblink.WRAPPER != 0) { p = liblink.Appendp(ctxt, p) p = load_g_cx(ctxt, p) // load g into CX } if !(cursym.Text.From.Scale&liblink.NOSPLIT != 0) { p = stacksplit(ctxt, p, autoffset, int32(textarg), bool2int(!(cursym.Text.From.Scale&liblink.NEEDCTXT != 0)), &q) // emit split check } if autoffset != 0 { if autoffset%int32(ctxt.Arch.Regsize) != 0 { ctxt.Diag("unaligned stack size %d", autoffset) } p = liblink.Appendp(ctxt, p) p.As = AADJSP p.From.Type_ = D_CONST p.From.Offset = int64(autoffset) p.Spadj = autoffset } else { // zero-byte stack adjustment. // Insert a fake non-zero adjustment so that stkcheck can // recognize the end of the stack-splitting prolog. p = liblink.Appendp(ctxt, p) p.As = ANOP p.Spadj = int32(-ctxt.Arch.Ptrsize) p = liblink.Appendp(ctxt, p) p.As = ANOP p.Spadj = int32(ctxt.Arch.Ptrsize) } if q != nil { q.Pcond = p } deltasp = autoffset if cursym.Text.From.Scale&liblink.WRAPPER != 0 { // if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame // // MOVQ g_panic(CX), BX // TESTQ BX, BX // JEQ end // LEAQ (autoffset+8)(SP), DI // CMPQ panic_argp(BX), DI // JNE end // MOVQ SP, panic_argp(BX) // end: // NOP // // The NOP is needed to give the jumps somewhere to land. // It is a liblink NOP, not an x86 NOP: it encodes to 0 instruction bytes. p = liblink.Appendp(ctxt, p) p.As = AMOVQ p.From.Type_ = D_INDIR + D_CX p.From.Offset = 4 * int64(ctxt.Arch.Ptrsize) // G.panic p.To.Type_ = D_BX if ctxt.Headtype == liblink.Hnacl { p.As = AMOVL p.From.Type_ = D_INDIR + D_R15 p.From.Scale = 1 p.From.Index = D_CX } p = liblink.Appendp(ctxt, p) p.As = ATESTQ p.From.Type_ = D_BX p.To.Type_ = D_BX if ctxt.Headtype == liblink.Hnacl { p.As = ATESTL } p = liblink.Appendp(ctxt, p) p.As = AJEQ p.To.Type_ = D_BRANCH p1 = p p = liblink.Appendp(ctxt, p) p.As = ALEAQ p.From.Type_ = D_INDIR + D_SP p.From.Offset = int64(autoffset) + 8 p.To.Type_ = D_DI if ctxt.Headtype == liblink.Hnacl { p.As = ALEAL } p = liblink.Appendp(ctxt, p) p.As = ACMPQ p.From.Type_ = D_INDIR + D_BX p.From.Offset = 0 // Panic.argp p.To.Type_ = D_DI if ctxt.Headtype == liblink.Hnacl { p.As = ACMPL p.From.Type_ = D_INDIR + D_R15 p.From.Scale = 1 p.From.Index = D_BX } p = liblink.Appendp(ctxt, p) p.As = AJNE p.To.Type_ = D_BRANCH p2 = p p = liblink.Appendp(ctxt, p) p.As = AMOVQ p.From.Type_ = D_SP p.To.Type_ = D_INDIR + D_BX p.To.Offset = 0 // Panic.argp if ctxt.Headtype == liblink.Hnacl { p.As = AMOVL p.To.Type_ = D_INDIR + D_R15 p.To.Scale = 1 p.To.Index = D_BX } p = liblink.Appendp(ctxt, p) p.As = ANOP p1.Pcond = p p2.Pcond = p } if ctxt.Debugzerostack != 0 && autoffset != 0 && !(cursym.Text.From.Scale&liblink.NOSPLIT != 0) { // 6l -Z means zero the stack frame on entry. // This slows down function calls but can help avoid // false positives in garbage collection. p = liblink.Appendp(ctxt, p) p.As = AMOVQ p.From.Type_ = D_SP p.To.Type_ = D_DI p = liblink.Appendp(ctxt, p) p.As = AMOVQ p.From.Type_ = D_CONST p.From.Offset = int64(autoffset) / 8 p.To.Type_ = D_CX p = liblink.Appendp(ctxt, p) p.As = AMOVQ p.From.Type_ = D_CONST p.From.Offset = 0 p.To.Type_ = D_AX p = liblink.Appendp(ctxt, p) p.As = AREP p = liblink.Appendp(ctxt, p) p.As = ASTOSQ } for ; p != nil; p = p.Link { pcsize = int(p.Mode) / 8 a = int(p.From.Type_) if a == D_AUTO { p.From.Offset += int64(deltasp) } if a == D_PARAM { p.From.Offset += int64(deltasp) + int64(pcsize) } a = int(p.To.Type_) if a == D_AUTO { p.To.Offset += int64(deltasp) } if a == D_PARAM { p.To.Offset += int64(deltasp) + int64(pcsize) } switch p.As { default: continue case APUSHL, APUSHFL: deltasp += 4 p.Spadj = 4 continue case APUSHQ, APUSHFQ: deltasp += 8 p.Spadj = 8 continue case APUSHW, APUSHFW: deltasp += 2 p.Spadj = 2 continue case APOPL, APOPFL: deltasp -= 4 p.Spadj = -4 continue case APOPQ, APOPFQ: deltasp -= 8 p.Spadj = -8 continue case APOPW, APOPFW: deltasp -= 2 p.Spadj = -2 continue case ARET: break } if autoffset != deltasp { ctxt.Diag("unbalanced PUSH/POP") } if autoffset != 0 { p.As = AADJSP p.From.Type_ = D_CONST p.From.Offset = int64(-autoffset) p.Spadj = -autoffset p = liblink.Appendp(ctxt, 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 != nil { // retjmp p.As = AJMP } } } func indir_cx(ctxt *liblink.Link, a *liblink.Addr) { if ctxt.Headtype == liblink.Hnacl { a.Type_ = D_INDIR + D_R15 a.Index = D_CX a.Scale = 1 return } a.Type_ = D_INDIR + D_CX } // Append code to p to load g into cx. // Overwrites p with the first instruction (no first appendp). // Overwriting p is unusual but it lets use this in both the // prologue (caller must call appendp first) and in the epilogue. // Returns last new instruction. func load_g_cx(ctxt *liblink.Link, p *liblink.Prog) *liblink.Prog { var next *liblink.Prog p.As = AMOVQ if ctxt.Arch.Ptrsize == 4 { p.As = AMOVL } p.From.Type_ = D_INDIR + D_TLS p.From.Offset = 0 p.To.Type_ = D_CX next = p.Link progedit(ctxt, p) for p.Link != next { p = p.Link } if p.From.Index == D_TLS { p.From.Scale = 2 } return p } // Append code to p to check for stack split. // Appends to (does not overwrite) p. // Assumes g is in CX. // Returns last new instruction. // On return, *jmpok is the instruction that should jump // to the stack frame allocation if no split is needed. func stacksplit(ctxt *liblink.Link, p *liblink.Prog, framesize int32, textarg int32, noctxt int, jmpok **liblink.Prog) *liblink.Prog { var q *liblink.Prog var q1 *liblink.Prog var cmp int var lea int var mov int var sub int cmp = ACMPQ lea = ALEAQ mov = AMOVQ sub = ASUBQ if ctxt.Headtype == liblink.Hnacl { cmp = ACMPL lea = ALEAL mov = AMOVL sub = ASUBL } q1 = nil if framesize <= liblink.StackSmall { // small stack: SP <= stackguard // CMPQ SP, stackguard p = liblink.Appendp(ctxt, p) p.As = int16(cmp) p.From.Type_ = D_SP indir_cx(ctxt, &p.To) p.To.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0 if ctxt.Cursym.Cfunc != 0 { p.To.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1 } } else if framesize <= liblink.StackBig { // large stack: SP-framesize <= stackguard-StackSmall // LEAQ -xxx(SP), AX // CMPQ AX, stackguard p = liblink.Appendp(ctxt, p) p.As = int16(lea) p.From.Type_ = D_INDIR + D_SP p.From.Offset = -(int64(framesize) - liblink.StackSmall) p.To.Type_ = D_AX p = liblink.Appendp(ctxt, p) p.As = int16(cmp) p.From.Type_ = D_AX indir_cx(ctxt, &p.To) p.To.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0 if ctxt.Cursym.Cfunc != 0 { p.To.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1 } } 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, $(framesize+(StackGuard-StackSmall)) p = liblink.Appendp(ctxt, p) p.As = int16(mov) indir_cx(ctxt, &p.From) p.From.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0 if ctxt.Cursym.Cfunc != 0 { p.From.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1 } p.To.Type_ = D_SI p = liblink.Appendp(ctxt, p) p.As = int16(cmp) p.From.Type_ = D_SI p.To.Type_ = D_CONST p.To.Offset = liblink.StackPreempt p = liblink.Appendp(ctxt, p) p.As = AJEQ p.To.Type_ = D_BRANCH q1 = p p = liblink.Appendp(ctxt, p) p.As = int16(lea) p.From.Type_ = D_INDIR + D_SP p.From.Offset = liblink.StackGuard p.To.Type_ = D_AX p = liblink.Appendp(ctxt, p) p.As = int16(sub) p.From.Type_ = D_SI p.To.Type_ = D_AX p = liblink.Appendp(ctxt, p) p.As = int16(cmp) p.From.Type_ = D_AX p.To.Type_ = D_CONST p.To.Offset = int64(framesize) + (liblink.StackGuard - liblink.StackSmall) } // common p = liblink.Appendp(ctxt, p) p.As = AJHI p.To.Type_ = D_BRANCH q = p p = liblink.Appendp(ctxt, p) p.As = ACALL p.To.Type_ = D_BRANCH if ctxt.Cursym.Cfunc != 0 { p.To.Sym = liblink.Linklookup(ctxt, "runtime.morestackc", 0) } else { p.To.Sym = ctxt.Symmorestack[noctxt] } p = liblink.Appendp(ctxt, p) p.As = AJMP p.To.Type_ = D_BRANCH p.Pcond = ctxt.Cursym.Text.Link if q != nil { q.Pcond = p.Link } if q1 != nil { q1.Pcond = q.Link } *jmpok = q return p } func follow(ctxt *liblink.Link, s *liblink.LSym) { var firstp *liblink.Prog var lastp *liblink.Prog ctxt.Cursym = s firstp = ctxt.Arch.Prg() lastp = firstp xfol(ctxt, s.Text, &lastp) lastp.Link = nil s.Text = firstp.Link } func nofollow(a int) int { switch a { case AJMP, ARET, AIRETL, AIRETQ, AIRETW, ARETFL, ARETFQ, ARETFW, AUNDEF: return 1 } return 0 } func pushpop(a int) int { switch a { case APUSHL, APUSHFL, APUSHQ, APUSHFQ, APUSHW, APUSHFW, APOPL, APOPFL, APOPQ, APOPFQ, APOPW, APOPFW: return 1 } return 0 } func relinv(a int) int { switch a { case AJEQ: return AJNE case AJNE: return AJEQ case AJLE: return AJGT case AJLS: return AJHI case AJLT: return AJGE case AJMI: return AJPL case AJGE: return AJLT case AJPL: return AJMI case AJGT: return AJLE case AJHI: return AJLS case AJCS: return AJCC case AJCC: return AJCS case AJPS: return AJPC case AJPC: return AJPS case AJOS: return AJOC case AJOC: return AJOS } log.Fatalf("unknown relation: %s", anames6[a]) return 0 } func xfol(ctxt *liblink.Link, p *liblink.Prog, last **liblink.Prog) { var q *liblink.Prog var i int var a int loop: if p == nil { return } if p.As == AJMP { q = p.Pcond if q != nil && q.As != ATEXT { /* mark instruction as done and continue layout at target of jump */ p.Mark = 1 p = q if p.Mark == 0 { goto loop } } } if p.Mark != 0 { /* * p goes here, but already used it elsewhere. * copy up to 4 instructions or else branch to other copy. */ i = 0 q = p for ; i < 4; (func() { i++; q = q.Link })() { if q == nil { break } if q == *last { break } a = int(q.As) if a == ANOP { i-- continue } if nofollow(a) != 0 || pushpop(a) != 0 { break // NOTE(rsc): arm does goto copy } if q.Pcond == nil || q.Pcond.Mark != 0 { continue } if a == ACALL || a == ALOOP { continue } for { if p.As == ANOP { p = p.Link continue } q = liblink.Copyp(ctxt, p) p = p.Link q.Mark = 1 (*last).Link = q *last = q if int(q.As) != a || q.Pcond == nil || q.Pcond.Mark != 0 { continue } q.As = int16(relinv(int(q.As))) p = q.Pcond q.Pcond = q.Link q.Link = p xfol(ctxt, q.Link, last) p = q.Link if p.Mark != 0 { return } goto loop /* */ } } q = ctxt.Arch.Prg() q.As = AJMP q.Lineno = p.Lineno q.To.Type_ = D_BRANCH q.To.Offset = p.Pc q.Pcond = p p = q } /* emit p */ p.Mark = 1 (*last).Link = p *last = p a = int(p.As) /* continue loop with what comes after p */ if nofollow(a) != 0 { return } if p.Pcond != nil && a != ACALL { /* * some kind of conditional branch. * recurse to follow one path. * continue loop on the other. */ q = liblink.Brchain(ctxt, p.Pcond) if q != nil { p.Pcond = q } q = liblink.Brchain(ctxt, p.Link) if q != nil { p.Link = q } if p.From.Type_ == D_CONST { if p.From.Offset == 1 { /* * expect conditional jump to be taken. * rewrite so that's the fall-through case. */ p.As = int16(relinv(a)) q = p.Link p.Link = p.Pcond p.Pcond = q } } else { q = p.Link if q.Mark != 0 { if a != ALOOP { p.As = int16(relinv(a)) p.Link = p.Pcond p.Pcond = q } } } xfol(ctxt, p.Link, last) if p.Pcond.Mark != 0 { return } p = p.Pcond goto loop } p = p.Link goto loop } func prg() *liblink.Prog { var p *liblink.Prog p = new(liblink.Prog) *p = zprg return p } var Linkamd64 = liblink.LinkArch{ ByteOrder: binary.LittleEndian, Pconv: Pconv, Name: "amd64", Thechar: '6', Endian: liblink.LittleEndian, Addstacksplit: addstacksplit, Assemble: span6, Datasize: datasize, Follow: follow, Iscall: iscall, Isdata: isdata, Prg: prg, Progedit: progedit, Settextflag: settextflag, Symtype: symtype, Textflag: textflag, Minlc: 1, Ptrsize: 8, Regsize: 8, D_ADDR: D_ADDR, D_AUTO: D_AUTO, D_BRANCH: D_BRANCH, D_CONST: D_CONST, D_EXTERN: D_EXTERN, D_FCONST: D_FCONST, D_NONE: D_NONE, D_PARAM: D_PARAM, D_SCONST: D_SCONST, D_STATIC: D_STATIC, ACALL: ACALL, ADATA: ADATA, AEND: AEND, AFUNCDATA: AFUNCDATA, AGLOBL: AGLOBL, AJMP: AJMP, ANOP: ANOP, APCDATA: APCDATA, ARET: ARET, ATEXT: ATEXT, ATYPE: ATYPE, AUSEFIELD: AUSEFIELD, } var Linkamd64p32 = liblink.LinkArch{ ByteOrder: binary.LittleEndian, Pconv: Pconv, Name: "amd64p32", Thechar: '6', Endian: liblink.LittleEndian, Addstacksplit: addstacksplit, Assemble: span6, Datasize: datasize, Follow: follow, Iscall: iscall, Isdata: isdata, Prg: prg, Progedit: progedit, Settextflag: settextflag, Symtype: symtype, Textflag: textflag, Minlc: 1, Ptrsize: 4, Regsize: 8, D_ADDR: D_ADDR, D_AUTO: D_AUTO, D_BRANCH: D_BRANCH, D_CONST: D_CONST, D_EXTERN: D_EXTERN, D_FCONST: D_FCONST, D_NONE: D_NONE, D_PARAM: D_PARAM, D_SCONST: D_SCONST, D_STATIC: D_STATIC, ACALL: ACALL, ADATA: ADATA, AEND: AEND, AFUNCDATA: AFUNCDATA, AGLOBL: AGLOBL, AJMP: AJMP, ANOP: ANOP, APCDATA: APCDATA, ARET: ARET, ATEXT: ATEXT, ATYPE: ATYPE, AUSEFIELD: AUSEFIELD, }