obj.go

// Copyright © 2015 The Go Authors.  All rights reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.

package riscv

import (
	"cmd/internal/obj"
	"cmd/internal/objabi"
	"cmd/internal/sys"
	"fmt"
	"log"
)

func buildop(ctxt *obj.Link) {}

// jalrToSym replaces p with a set of Progs needed to jump to the Sym in p.
// lr is the link register to use for the JALR.
// p must be a CALL, JMP or RET.
func jalrToSym(ctxt *obj.Link, p *obj.Prog, newprog obj.ProgAlloc, lr int16) *obj.Prog {
	if p.As != obj.ACALL && p.As != obj.AJMP && p.As != obj.ARET && p.As != obj.ADUFFZERO && p.As != obj.ADUFFCOPY {
		ctxt.Diag("unexpected Prog in jalrToSym: %v", p)
		return p
	}

	// TODO(jsing): Consider using a single JAL instruction and teaching
	// the linker to provide trampolines for the case where the destination
	// offset is too large. This would potentially reduce instructions for
	// the common case, but would require three instructions to go via the
	// trampoline.

	to := p.To

	p.As = AAUIPC
	p.Mark |= NEED_PCREL_ITYPE_RELOC
	p.SetFrom3(obj.Addr{Type: obj.TYPE_CONST, Offset: to.Offset, Sym: to.Sym})
	p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: 0}
	p.Reg = 0
	p.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}
	p = obj.Appendp(p, newprog)

	// Leave Sym only for the CALL reloc in assemble.
	p.As = AJALR
	p.From.Type = obj.TYPE_REG
	p.From.Reg = lr
	p.Reg = 0
	p.To.Type = obj.TYPE_REG
	p.To.Reg = REG_TMP
	p.To.Sym = to.Sym

	return p
}

// progedit is called individually for each *obj.Prog. It normalizes instruction
// formats and eliminates as many pseudo-instructions as possible.
func progedit(ctxt *obj.Link, p *obj.Prog, newprog obj.ProgAlloc) {

	// Expand binary instructions to ternary ones.
	if p.Reg == 0 {
		switch p.As {
		case AADDI, ASLTI, ASLTIU, AANDI, AORI, AXORI, ASLLI, ASRLI, ASRAI,
			AADD, AAND, AOR, AXOR, ASLL, ASRL, ASUB, ASRA,
			AMUL, AMULH, AMULHU, AMULHSU, AMULW, ADIV, ADIVU, ADIVW, ADIVUW,
			AREM, AREMU, AREMW, AREMUW:
			p.Reg = p.To.Reg
		}
	}

	// Rewrite instructions with constant operands to refer to the immediate
	// form of the instruction.
	if p.From.Type == obj.TYPE_CONST {
		switch p.As {
		case AADD:
			p.As = AADDI
		case ASLT:
			p.As = ASLTI
		case ASLTU:
			p.As = ASLTIU
		case AAND:
			p.As = AANDI
		case AOR:
			p.As = AORI
		case AXOR:
			p.As = AXORI
		case ASLL:
			p.As = ASLLI
		case ASRL:
			p.As = ASRLI
		case ASRA:
			p.As = ASRAI
		}
	}

	switch p.As {
	case obj.AJMP:
		// Turn JMP into JAL ZERO or JALR ZERO.
		p.From.Type = obj.TYPE_REG
		p.From.Reg = REG_ZERO

		switch p.To.Type {
		case obj.TYPE_BRANCH:
			p.As = AJAL
		case obj.TYPE_MEM:
			switch p.To.Name {
			case obj.NAME_NONE:
				p.As = AJALR
			case obj.NAME_EXTERN, obj.NAME_STATIC:
				// Handled in preprocess.
			default:
				ctxt.Diag("unsupported name %d for %v", p.To.Name, p)
			}
		default:
			panic(fmt.Sprintf("unhandled type %+v", p.To.Type))
		}

	case obj.ACALL:
		switch p.To.Type {
		case obj.TYPE_MEM:
			// Handled in preprocess.
		case obj.TYPE_REG:
			p.As = AJALR
			p.From.Type = obj.TYPE_REG
			p.From.Reg = REG_LR
		default:
			ctxt.Diag("unknown destination type %+v in CALL: %v", p.To.Type, p)
		}

	case obj.AUNDEF:
		p.As = AEBREAK

	case ASCALL:
		// SCALL is the old name for ECALL.
		p.As = AECALL

	case ASBREAK:
		// SBREAK is the old name for EBREAK.
		p.As = AEBREAK

	case AMOV:
		// Put >32-bit constants in memory and load them.
		if p.From.Type == obj.TYPE_CONST && p.From.Name == obj.NAME_NONE && p.From.Reg == 0 && int64(int32(p.From.Offset)) != p.From.Offset {
			p.From.Type = obj.TYPE_MEM
			p.From.Sym = ctxt.Int64Sym(p.From.Offset)
			p.From.Name = obj.NAME_EXTERN
			p.From.Offset = 0
		}
	}
}

// addrToReg extracts the register from an Addr, handling special Addr.Names.
func addrToReg(a obj.Addr) int16 {
	switch a.Name {
	case obj.NAME_PARAM, obj.NAME_AUTO:
		return REG_SP
	}
	return a.Reg
}

// movToLoad converts a MOV mnemonic into the corresponding load instruction.
func movToLoad(mnemonic obj.As) obj.As {
	switch mnemonic {
	case AMOV:
		return ALD
	case AMOVB:
		return ALB
	case AMOVH:
		return ALH
	case AMOVW:
		return ALW
	case AMOVBU:
		return ALBU
	case AMOVHU:
		return ALHU
	case AMOVWU:
		return ALWU
	case AMOVF:
		return AFLW
	case AMOVD:
		return AFLD
	default:
		panic(fmt.Sprintf("%+v is not a MOV", mnemonic))
	}
}

// movToStore converts a MOV mnemonic into the corresponding store instruction.
func movToStore(mnemonic obj.As) obj.As {
	switch mnemonic {
	case AMOV:
		return ASD
	case AMOVB:
		return ASB
	case AMOVH:
		return ASH
	case AMOVW:
		return ASW
	case AMOVF:
		return AFSW
	case AMOVD:
		return AFSD
	default:
		panic(fmt.Sprintf("%+v is not a MOV", mnemonic))
	}
}

// rewriteMOV rewrites MOV pseudo-instructions.
func rewriteMOV(ctxt *obj.Link, newprog obj.ProgAlloc, p *obj.Prog) {
	switch p.As {
	case AMOV, AMOVB, AMOVH, AMOVW, AMOVBU, AMOVHU, AMOVWU, AMOVF, AMOVD:
	default:
		panic(fmt.Sprintf("%+v is not a MOV pseudo-instruction", p.As))
	}

	switch p.From.Type {
	case obj.TYPE_MEM: // MOV c(Rs), Rd -> L $c, Rs, Rd
		switch p.From.Name {
		case obj.NAME_AUTO, obj.NAME_PARAM, obj.NAME_NONE:
			if p.To.Type != obj.TYPE_REG {
				ctxt.Diag("unsupported load at %v", p)
			}
			p.As = movToLoad(p.As)
			p.From.Reg = addrToReg(p.From)

		case obj.NAME_EXTERN, obj.NAME_STATIC:
			// AUIPC $off_hi, R
			// L $off_lo, R
			as := p.As
			to := p.To

			p.As = AAUIPC
			p.Mark |= NEED_PCREL_ITYPE_RELOC
			p.SetFrom3(obj.Addr{Type: obj.TYPE_CONST, Offset: p.From.Offset, Sym: p.From.Sym})
			p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: 0}
			p.Reg = 0
			p.To = obj.Addr{Type: obj.TYPE_REG, Reg: to.Reg}
			p = obj.Appendp(p, newprog)

			p.As = movToLoad(as)
			p.From = obj.Addr{Type: obj.TYPE_MEM, Reg: to.Reg, Offset: 0}
			p.To = to

		default:
			ctxt.Diag("unsupported name %d for %v", p.From.Name, p)
		}

	case obj.TYPE_REG:
		switch p.To.Type {
		case obj.TYPE_REG:
			switch p.As {
			case AMOV, AMOVB, AMOVH, AMOVW, AMOVBU, AMOVHU, AMOVWU, AMOVF, AMOVD:
			default:
				ctxt.Diag("unsupported register-register move at %v", p)
			}

		case obj.TYPE_MEM: // MOV Rs, c(Rd) -> S $c, Rs, Rd
			switch p.As {
			case AMOVBU, AMOVHU, AMOVWU:
				ctxt.Diag("unsupported unsigned store at %v", p)
				return
			}
			switch p.To.Name {
			case obj.NAME_AUTO, obj.NAME_PARAM, obj.NAME_NONE:
				p.As = movToStore(p.As)
				p.To.Reg = addrToReg(p.To)

			case obj.NAME_EXTERN, obj.NAME_STATIC:
				// AUIPC $off_hi, TMP
				// S $off_lo, TMP, R
				as := p.As
				from := p.From

				p.As = AAUIPC
				p.Mark |= NEED_PCREL_STYPE_RELOC
				p.SetFrom3(obj.Addr{Type: obj.TYPE_CONST, Offset: p.To.Offset, Sym: p.To.Sym})
				p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: 0}
				p.Reg = 0
				p.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}
				p = obj.Appendp(p, newprog)

				p.As = movToStore(as)
				p.From = from
				p.To = obj.Addr{Type: obj.TYPE_MEM, Reg: REG_TMP, Offset: 0}

			default:
				ctxt.Diag("unsupported name %d for %v", p.From.Name, p)
			}

		default:
			ctxt.Diag("unsupported MOV at %v", p)
		}

	case obj.TYPE_CONST:
		// MOV $c, R
		// If c is small enough, convert to:
		//   ADD $c, ZERO, R
		// If not, convert to:
		//   LUI top20bits(c), R
		//   ADD bottom12bits(c), R, R
		if p.As != AMOV {
			ctxt.Diag("%v: unsupported constant load", p)
		}
		if p.To.Type != obj.TYPE_REG {
			ctxt.Diag("%v: constant load must target register", p)
		}
		off := p.From.Offset
		to := p.To

		low, high, err := Split32BitImmediate(off)
		if err != nil {
			ctxt.Diag("%v: constant %d too large: %v", p, off, err)
		}

		// LUI is only necessary if the offset doesn't fit in 12-bits.
		needLUI := high != 0
		if needLUI {
			p.As = ALUI
			p.To = to
			// Pass top 20 bits to LUI.
			p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: high}
			p = obj.Appendp(p, newprog)
		}
		p.As = AADDIW
		p.To = to
		p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: low}
		p.Reg = REG_ZERO
		if needLUI {
			p.Reg = to.Reg
		}

	case obj.TYPE_ADDR: // MOV $sym+off(SP/SB), R
		if p.To.Type != obj.TYPE_REG || p.As != AMOV {
			ctxt.Diag("unsupported addr MOV at %v", p)
		}
		switch p.From.Name {
		case obj.NAME_EXTERN, obj.NAME_STATIC:
			// AUIPC $off_hi, R
			// ADDI $off_lo, R
			to := p.To

			p.As = AAUIPC
			p.Mark |= NEED_PCREL_ITYPE_RELOC
			p.SetFrom3(obj.Addr{Type: obj.TYPE_CONST, Offset: p.From.Offset, Sym: p.From.Sym})
			p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: 0}
			p.Reg = 0
			p.To = to
			p = obj.Appendp(p, newprog)

			p.As = AADDI
			p.From = obj.Addr{Type: obj.TYPE_CONST}
			p.Reg = to.Reg
			p.To = to

		case obj.NAME_PARAM, obj.NAME_AUTO:
			p.As = AADDI
			p.Reg = REG_SP
			p.From.Type = obj.TYPE_CONST

		case obj.NAME_NONE:
			p.As = AADDI
			p.Reg = p.From.Reg
			p.From.Type = obj.TYPE_CONST
			p.From.Reg = 0

		default:
			ctxt.Diag("bad addr MOV from name %v at %v", p.From.Name, p)
		}

	default:
		ctxt.Diag("unsupported MOV at %v", p)
	}
}

// InvertBranch inverts the condition of a conditional branch.
func InvertBranch(as obj.As) obj.As {
	switch as {
	case ABEQ:
		return ABNE
	case ABEQZ:
		return ABNEZ
	case ABGE:
		return ABLT
	case ABGEU:
		return ABLTU
	case ABGEZ:
		return ABLTZ
	case ABGT:
		return ABLE
	case ABGTU:
		return ABLEU
	case ABGTZ:
		return ABLEZ
	case ABLE:
		return ABGT
	case ABLEU:
		return ABGTU
	case ABLEZ:
		return ABGTZ
	case ABLT:
		return ABGE
	case ABLTU:
		return ABGEU
	case ABLTZ:
		return ABGEZ
	case ABNE:
		return ABEQ
	case ABNEZ:
		return ABEQZ
	default:
		panic("InvertBranch: not a branch")
	}
}

// containsCall reports whether the symbol contains a CALL (or equivalent)
// instruction. Must be called after progedit.
func containsCall(sym *obj.LSym) bool {
	// CALLs are CALL or JAL(R) with link register LR.
	for p := sym.Func().Text; p != nil; p = p.Link {
		switch p.As {
		case obj.ACALL, obj.ADUFFZERO, obj.ADUFFCOPY:
			return true
		case AJAL, AJALR:
			if p.From.Type == obj.TYPE_REG && p.From.Reg == REG_LR {
				return true
			}
		}
	}

	return false
}

// setPCs sets the Pc field in all instructions reachable from p.
// It uses pc as the initial value.
func setPCs(p *obj.Prog, pc int64) {
	for ; p != nil; p = p.Link {
		p.Pc = pc
		for _, ins := range instructionsForProg(p) {
			pc += int64(ins.length())
		}
	}
}

// stackOffset updates Addr offsets based on the current stack size.
//
// The stack looks like:
// -------------------
// |                 |
// |      PARAMs     |
// |                 |
// |                 |
// -------------------
// |    Parent RA    |   SP on function entry
// -------------------
// |                 |
// |                 |
// |       AUTOs     |
// |                 |
// |                 |
// -------------------
// |        RA       |   SP during function execution
// -------------------
//
// FixedFrameSize makes other packages aware of the space allocated for RA.
//
// A nicer version of this diagram can be found on slide 21 of the presentation
// attached to:
//
//   https://golang.org/issue/16922#issuecomment-243748180
//
func stackOffset(a *obj.Addr, stacksize int64) {
	switch a.Name {
	case obj.NAME_AUTO:
		// Adjust to the top of AUTOs.
		a.Offset += stacksize
	case obj.NAME_PARAM:
		// Adjust to the bottom of PARAMs.
		a.Offset += stacksize + 8
	}
}

// preprocess generates prologue and epilogue code, computes PC-relative branch
// and jump offsets, and resolves pseudo-registers.
//
// preprocess is called once per linker symbol.
//
// When preprocess finishes, all instructions in the symbol are either
// concrete, real RISC-V instructions or directive pseudo-ops like TEXT,
// PCDATA, and FUNCDATA.
func preprocess(ctxt *obj.Link, cursym *obj.LSym, newprog obj.ProgAlloc) {
	if cursym.Func().Text == nil || cursym.Func().Text.Link == nil {
		return
	}

	// Generate the prologue.
	text := cursym.Func().Text
	if text.As != obj.ATEXT {
		ctxt.Diag("preprocess: found symbol that does not start with TEXT directive")
		return
	}

	stacksize := text.To.Offset
	if stacksize == -8 {
		// Historical way to mark NOFRAME.
		text.From.Sym.Set(obj.AttrNoFrame, true)
		stacksize = 0
	}
	if stacksize < 0 {
		ctxt.Diag("negative frame size %d - did you mean NOFRAME?", stacksize)
	}
	if text.From.Sym.NoFrame() {
		if stacksize != 0 {
			ctxt.Diag("NOFRAME functions must have a frame size of 0, not %d", stacksize)
		}
	}

	if !containsCall(cursym) {
		text.From.Sym.Set(obj.AttrLeaf, true)
		if stacksize == 0 {
			// A leaf function with no locals has no frame.
			text.From.Sym.Set(obj.AttrNoFrame, true)
		}
	}

	// Save LR unless there is no frame.
	if !text.From.Sym.NoFrame() {
		stacksize += ctxt.FixedFrameSize()
	}

	cursym.Func().Args = text.To.Val.(int32)
	cursym.Func().Locals = int32(stacksize)

	prologue := text

	if !cursym.Func().Text.From.Sym.NoSplit() {
		prologue = stacksplit(ctxt, prologue, cursym, newprog, stacksize) // emit split check
	}

	if stacksize != 0 {
		prologue = ctxt.StartUnsafePoint(prologue, newprog)

		// Actually save LR.
		prologue = obj.Appendp(prologue, newprog)
		prologue.As = AMOV
		prologue.From = obj.Addr{Type: obj.TYPE_REG, Reg: REG_LR}
		prologue.To = obj.Addr{Type: obj.TYPE_MEM, Reg: REG_SP, Offset: -stacksize}

		// Insert stack adjustment.
		prologue = obj.Appendp(prologue, newprog)
		prologue.As = AADDI
		prologue.From = obj.Addr{Type: obj.TYPE_CONST, Offset: -stacksize}
		prologue.Reg = REG_SP
		prologue.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_SP}
		prologue.Spadj = int32(stacksize)

		prologue = ctxt.EndUnsafePoint(prologue, newprog, -1)
	}

	if cursym.Func().Text.From.Sym.Wrapper() {
		// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
		//
		//   MOV g_panic(g), X11
		//   BNE X11, ZERO, adjust
		// end:
		//   NOP
		// ...rest of function..
		// adjust:
		//   MOV panic_argp(X11), X12
		//   ADD $(autosize+FIXED_FRAME), SP, X13
		//   BNE X12, X13, end
		//   ADD $FIXED_FRAME, SP, X12
		//   MOV X12, panic_argp(X11)
		//   JMP end
		//
		// The NOP is needed to give the jumps somewhere to land.

		ldpanic := obj.Appendp(prologue, newprog)

		ldpanic.As = AMOV
		ldpanic.From = obj.Addr{Type: obj.TYPE_MEM, Reg: REGG, Offset: 4 * int64(ctxt.Arch.PtrSize)} // G.panic
		ldpanic.Reg = 0
		ldpanic.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_X11}

		bneadj := obj.Appendp(ldpanic, newprog)
		bneadj.As = ABNE
		bneadj.From = obj.Addr{Type: obj.TYPE_REG, Reg: REG_X11}
		bneadj.Reg = REG_ZERO
		bneadj.To.Type = obj.TYPE_BRANCH

		endadj := obj.Appendp(bneadj, newprog)
		endadj.As = obj.ANOP

		last := endadj
		for last.Link != nil {
			last = last.Link
		}

		getargp := obj.Appendp(last, newprog)
		getargp.As = AMOV
		getargp.From = obj.Addr{Type: obj.TYPE_MEM, Reg: REG_X11, Offset: 0} // Panic.argp
		getargp.Reg = 0
		getargp.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_X12}

		bneadj.To.SetTarget(getargp)

		calcargp := obj.Appendp(getargp, newprog)
		calcargp.As = AADDI
		calcargp.From = obj.Addr{Type: obj.TYPE_CONST, Offset: stacksize + ctxt.FixedFrameSize()}
		calcargp.Reg = REG_SP
		calcargp.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_X13}

		testargp := obj.Appendp(calcargp, newprog)
		testargp.As = ABNE
		testargp.From = obj.Addr{Type: obj.TYPE_REG, Reg: REG_X12}
		testargp.Reg = REG_X13
		testargp.To.Type = obj.TYPE_BRANCH
		testargp.To.SetTarget(endadj)

		adjargp := obj.Appendp(testargp, newprog)
		adjargp.As = AADDI
		adjargp.From = obj.Addr{Type: obj.TYPE_CONST, Offset: int64(ctxt.Arch.PtrSize)}
		adjargp.Reg = REG_SP
		adjargp.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_X12}

		setargp := obj.Appendp(adjargp, newprog)
		setargp.As = AMOV
		setargp.From = obj.Addr{Type: obj.TYPE_REG, Reg: REG_X12}
		setargp.Reg = 0
		setargp.To = obj.Addr{Type: obj.TYPE_MEM, Reg: REG_X11, Offset: 0} // Panic.argp

		godone := obj.Appendp(setargp, newprog)
		godone.As = AJAL
		godone.From = obj.Addr{Type: obj.TYPE_REG, Reg: REG_ZERO}
		godone.To.Type = obj.TYPE_BRANCH
		godone.To.SetTarget(endadj)
	}

	// Update stack-based offsets.
	for p := cursym.Func().Text; p != nil; p = p.Link {
		stackOffset(&p.From, stacksize)
		stackOffset(&p.To, stacksize)
	}

	// Additional instruction rewriting.
	for p := cursym.Func().Text; p != nil; p = p.Link {
		switch p.As {
		case obj.AGETCALLERPC:
			if cursym.Leaf() {
				// MOV LR, Rd
				p.As = AMOV
				p.From.Type = obj.TYPE_REG
				p.From.Reg = REG_LR
			} else {
				// MOV (RSP), Rd
				p.As = AMOV
				p.From.Type = obj.TYPE_MEM
				p.From.Reg = REG_SP
			}

		case obj.ACALL, obj.ADUFFZERO, obj.ADUFFCOPY:
			switch p.To.Type {
			case obj.TYPE_MEM:
				jalrToSym(ctxt, p, newprog, REG_LR)
			}

		case obj.AJMP:
			switch p.To.Type {
			case obj.TYPE_MEM:
				switch p.To.Name {
				case obj.NAME_EXTERN, obj.NAME_STATIC:
					// JMP to symbol.
					jalrToSym(ctxt, p, newprog, REG_ZERO)
				}
			}

		case obj.ARET:
			// Replace RET with epilogue.
			retJMP := p.To.Sym

			if stacksize != 0 {
				// Restore LR.
				p.As = AMOV
				p.From = obj.Addr{Type: obj.TYPE_MEM, Reg: REG_SP, Offset: 0}
				p.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_LR}
				p = obj.Appendp(p, newprog)

				p.As = AADDI
				p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: stacksize}
				p.Reg = REG_SP
				p.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_SP}
				p.Spadj = int32(-stacksize)
				p = obj.Appendp(p, newprog)
			}

			if retJMP != nil {
				p.As = obj.ARET
				p.To.Sym = retJMP
				p = jalrToSym(ctxt, p, newprog, REG_ZERO)
			} else {
				p.As = AJALR
				p.From = obj.Addr{Type: obj.TYPE_REG, Reg: REG_ZERO}
				p.Reg = 0
				p.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_LR}
			}

			// "Add back" the stack removed in the previous instruction.
			//
			// This is to avoid confusing pctospadj, which sums
			// Spadj from function entry to each PC, and shouldn't
			// count adjustments from earlier epilogues, since they
			// won't affect later PCs.
			p.Spadj = int32(stacksize)

		case AADDI:
			// Refine Spadjs account for adjustment via ADDI instruction.
			if p.To.Type == obj.TYPE_REG && p.To.Reg == REG_SP && p.From.Type == obj.TYPE_CONST {
				p.Spadj = int32(-p.From.Offset)
			}
		}

		if p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP && p.Spadj == 0 {
			f := cursym.Func()
			if f.FuncFlag&objabi.FuncFlag_SPWRITE == 0 {
				f.FuncFlag |= objabi.FuncFlag_SPWRITE
				if ctxt.Debugvlog || !ctxt.IsAsm {
					ctxt.Logf("auto-SPWRITE: %s %v\n", cursym.Name, p)
					if !ctxt.IsAsm {
						ctxt.Diag("invalid auto-SPWRITE in non-assembly")
						ctxt.DiagFlush()
						log.Fatalf("bad SPWRITE")
					}
				}
			}
		}
	}

	// Rewrite MOV pseudo-instructions. This cannot be done in
	// progedit, as SP offsets need to be applied before we split
	// up some of the Addrs.
	for p := cursym.Func().Text; p != nil; p = p.Link {
		switch p.As {
		case AMOV, AMOVB, AMOVH, AMOVW, AMOVBU, AMOVHU, AMOVWU, AMOVF, AMOVD:
			rewriteMOV(ctxt, newprog, p)
		}
	}

	// Split immediates larger than 12-bits.
	for p := cursym.Func().Text; p != nil; p = p.Link {
		switch p.As {
		// <opi> $imm, REG, TO
		case AADDI, AANDI, AORI, AXORI:
			// LUI $high, TMP
			// ADDI $low, TMP, TMP
			// <op> TMP, REG, TO
			q := *p
			low, high, err := Split32BitImmediate(p.From.Offset)
			if err != nil {
				ctxt.Diag("%v: constant %d too large", p, p.From.Offset, err)
			}
			if high == 0 {
				break // no need to split
			}

			// Split into two additions if possible.
			imm := q.From.Offset
			const minInt12, maxInt12 = -(1 << 11), (1 << 11) - 1
			if q.As == AADDI && 2*minInt12 <= imm && imm <= 2*maxInt12 {
				imm0, imm1 := imm/2, imm-imm/2
				// ADDI $(imm/2), REG, TO
				p.Spadj = 0 // needed if TO is SP
				p.As = AADDI
				p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: imm0}
				p.Reg = q.Reg
				p.To = q.To
				p = obj.Appendp(p, newprog)
				// ADDI $(imm-imm/2), TO, TO
				p.Spadj = q.Spadj
				p.As = AADDI
				p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: imm1}
				p.Reg = q.To.Reg
				p.To = q.To
				break
			}

			p.As = ALUI
			p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: high}
			p.Reg = 0
			p.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}
			p.Spadj = 0 // needed if TO is SP
			p = obj.Appendp(p, newprog)

			p.As = AADDIW
			p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: low}
			p.Reg = REG_TMP
			p.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}
			p = obj.Appendp(p, newprog)

			switch q.As {
			case AADDI:
				p.As = AADD
			case AANDI:
				p.As = AAND
			case AORI:
				p.As = AOR
			case AXORI:
				p.As = AXOR
			default:
				ctxt.Diag("unsupported instruction %v for splitting", q)
			}
			p.Spadj = q.Spadj
			p.To = q.To
			p.Reg = q.Reg
			p.From = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}

		// <load> $imm, REG, TO (load $imm+(REG), TO)
		case ALD, ALB, ALH, ALW, ALBU, ALHU, ALWU, AFLW, AFLD:
			low, high, err := Split32BitImmediate(p.From.Offset)
			if err != nil {
				ctxt.Diag("%v: constant %d too large", p, p.From.Offset)
			}
			if high == 0 {
				break // no need to split
			}
			q := *p

			// LUI $high, TMP
			// ADD TMP, REG, TMP
			// <load> $low, TMP, TO
			p.As = ALUI
			p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: high}
			p.Reg = 0
			p.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}
			p.Spadj = 0 // needed if TO is SP
			p = obj.Appendp(p, newprog)

			p.As = AADD
			p.From = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}
			p.Reg = q.From.Reg
			p.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}
			p = obj.Appendp(p, newprog)

			p.As = q.As
			p.To = q.To
			p.From = obj.Addr{Type: obj.TYPE_MEM, Reg: REG_TMP, Offset: low}
			p.Reg = obj.REG_NONE

		// <store> $imm, REG, TO (store $imm+(TO), REG)
		case ASD, ASB, ASH, ASW, AFSW, AFSD:
			low, high, err := Split32BitImmediate(p.To.Offset)
			if err != nil {
				ctxt.Diag("%v: constant %d too large", p, p.To.Offset)
			}
			if high == 0 {
				break // no need to split
			}
			q := *p

			// LUI $high, TMP
			// ADD TMP, TO, TMP
			// <store> $low, REG, TMP
			p.As = ALUI
			p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: high}
			p.Reg = 0
			p.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}
			p.Spadj = 0 // needed if TO is SP
			p = obj.Appendp(p, newprog)

			p.As = AADD
			p.From = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}
			p.Reg = q.To.Reg
			p.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}
			p = obj.Appendp(p, newprog)

			p.As = q.As
			p.From = obj.Addr{Type: obj.TYPE_REG, Reg: q.From.Reg, Offset: 0}
			p.To = obj.Addr{Type: obj.TYPE_MEM, Reg: REG_TMP, Offset: low}
		}
	}

	// Compute instruction addresses.  Once we do that, we need to check for
	// overextended jumps and branches.  Within each iteration, Pc differences
	// are always lower bounds (since the program gets monotonically longer,
	// a fixed point will be reached).  No attempt to handle functions > 2GiB.
	for {
		rescan := false
		setPCs(cursym.Func().Text, 0)

		for p := cursym.Func().Text; p != nil; p = p.Link {
			switch p.As {
			case ABEQ, ABEQZ, ABGE, ABGEU, ABGEZ, ABGT, ABGTU, ABGTZ, ABLE, ABLEU, ABLEZ, ABLT, ABLTU, ABLTZ, ABNE, ABNEZ:
				if p.To.Type != obj.TYPE_BRANCH {
					panic("assemble: instruction with branch-like opcode lacks destination")
				}
				offset := p.To.Target().Pc - p.Pc
				if offset < -4096 || 4096 <= offset {
					// Branch is long.  Replace it with a jump.
					jmp := obj.Appendp(p, newprog)
					jmp.As = AJAL
					jmp.From = obj.Addr{Type: obj.TYPE_REG, Reg: REG_ZERO}
					jmp.To = obj.Addr{Type: obj.TYPE_BRANCH}
					jmp.To.SetTarget(p.To.Target())

					p.As = InvertBranch(p.As)
					p.To.SetTarget(jmp.Link)

					// We may have made previous branches too long,
					// so recheck them.
					rescan = true
				}
			case AJAL:
				if p.To.Target() == nil {
					panic("intersymbol jumps should be expressed as AUIPC+JALR")
				}
				offset := p.To.Target().Pc - p.Pc
				if offset < -(1<<20) || (1<<20) <= offset {
					// Replace with 2-instruction sequence. This assumes
					// that TMP is not live across J instructions, since
					// it is reserved by SSA.
					jmp := obj.Appendp(p, newprog)
					jmp.As = AJALR
					jmp.From = p.From
					jmp.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}

					// p.From is not generally valid, however will be
					// fixed up in the next loop.
					p.As = AAUIPC
					p.From = obj.Addr{Type: obj.TYPE_BRANCH, Sym: p.From.Sym}
					p.From.SetTarget(p.To.Target())
					p.Reg = 0
					p.To = obj.Addr{Type: obj.TYPE_REG, Reg: REG_TMP}

					rescan = true
				}
			}
		}

		if !rescan {
			break
		}
	}

	// Now that there are no long branches, resolve branch and jump targets.
	// At this point, instruction rewriting which changes the number of
	// instructions will break everything--don't do it!
	for p := cursym.Func().Text; p != nil; p = p.Link {
		switch p.As {
		case ABEQ, ABEQZ, ABGE, ABGEU, ABGEZ, ABGT, ABGTU, ABGTZ, ABLE, ABLEU, ABLEZ, ABLT, ABLTU, ABLTZ, ABNE, ABNEZ, AJAL:
			switch p.To.Type {
			case obj.TYPE_BRANCH:
				p.To.Type, p.To.Offset = obj.TYPE_CONST, p.To.Target().Pc-p.Pc
			case obj.TYPE_MEM:
				panic("unhandled type")
			}

		case AAUIPC:
			if p.From.Type == obj.TYPE_BRANCH {
				low, high, err := Split32BitImmediate(p.From.Target().Pc - p.Pc)
				if err != nil {
					ctxt.Diag("%v: jump displacement %d too large", p, p.To.Target().Pc-p.Pc)
				}
				p.From = obj.Addr{Type: obj.TYPE_CONST, Offset: high, Sym: cursym}
				p.Link.From.Offset = low
			}
		}
	}

	// Validate all instructions - this provides nice error messages.
	for p := cursym.Func().Text; p != nil; p = p.Link {
		for _, ins := range instructionsForProg(p) {
			ins.validate(ctxt)
		}
	}
}

func stacksplit(ctxt *obj.Link, p *obj.Prog, cursym *obj.LSym, newprog obj.ProgAlloc, framesize int64) *obj.Prog {
	// Leaf function with no frame is effectively NOSPLIT.
	if framesize == 0 {
		return p
	}

	// MOV	g_stackguard(g), X10
	p = obj.Appendp(p, newprog)
	p.As = AMOV
	p.From.Type = obj.TYPE_MEM
	p.From.Reg = REGG
	p.From.Offset = 2 * int64(ctxt.Arch.PtrSize) // G.stackguard0
	if cursym.CFunc() {