func (mach X86) opCheckDivideByZero(code gen.RegCoder, t types.T, reg regs.R) {
var end links.L
Test.opFromReg(code, t, reg, reg)
Jne.rel8.opStub(code)
end.AddSite(code.Len())
code.OpTrapCall(traps.IntegerDivideByZero)
end.Addr = code.Len()
mach.updateBranches8(code, &end)
}
func (mach X86) opConvertUnsignedI64ToFloat(code gen.Coder, resultType types.T, inputReg regs.R) {
// this algorithm is copied from code generated by gcc and clang:
var done links.L
var huge links.L
// TODO: allocate target reg
Test.opFromReg(code, types.I64, inputReg, inputReg)
Js.rel8.opStub(code)
huge.AddSite(code.Len())
// max. 63-bit value
Cvtsi2sSSE.opReg(code, resultType, types.I64, regResult, inputReg)
JmpRel.rel8.opStub(code)
done.AddSite(code.Len())
huge.Addr = code.Len()
mach.updateBranches8(code, &huge)
// 64-bit value
Mov.opFromReg(code, types.I64, regScratch, inputReg)
And.opImm(code, types.I64, regScratch, 1)
ShrImm.op(code, types.I64, inputReg, 1)
Or.opFromReg(code, types.I64, inputReg, regScratch)
Cvtsi2sSSE.opReg(code, resultType, types.I64, regResult, inputReg)
AddsSSE.opFromReg(code, resultType, regResult, regResult)
done.Addr = code.Len()
mach.updateBranches8(code, &done)
}
// OpCallIndirect using table index located in result register.
func (mach X86) OpCallIndirect(code gen.Coder, tableLen, sigIndex int32) int32 {
var outOfBounds links.L
var checksOut links.L
mach.opCompareBounds(code, regResult, tableLen)
Jle.rel8.opStub(code) // TODO: is this the correct comparison?
outOfBounds.AddSite(code.Len())
Mov.opFromAddr(code, types.I64, regResult, 3, regResult, code.RODataAddr()+gen.ROTableAddr)
Mov.opFromReg(code, types.I32, regScratch, regResult) // zero-extended function address
ShrImm.op(code, types.I64, regResult, 32) // signature index
Cmp.opImm(code, types.I32, regResult, sigIndex)
Je.rel8.opStub(code)
checksOut.AddSite(code.Len())
code.OpTrapCall(traps.IndirectCallSignature)
outOfBounds.Addr = code.Len()
mach.updateBranches8(code, &outOfBounds)
code.OpTrapCall(traps.IndirectCallIndex)
checksOut.Addr = code.Len()
mach.updateBranches8(code, &checksOut)
Add.opFromReg(code, types.I64, regScratch, regTextBase)
Call.opReg(code, regScratch)
return code.Len()
}
func (mach X86) OpTrapIfStackExhausted(code gen.Coder) (stackCheckAddr int32) {
var checked links.L
Lea.opFromStack(code, types.I64, regScratch, -0x80000000) // reserve 32-bit displacement
stackCheckAddr = code.Len()
Cmp.opFromReg(code, types.I64, regScratch, regStackLimit)
Jge.rel8.opStub(code)
checked.AddSite(code.Len())
code.OpTrapCall(traps.CallStackExhausted)
checked.Addr = code.Len()
mach.updateBranches8(code, &checked)
return
}
func (mach X86) OpInit(code gen.OpCoder, startAddr int32) (retAddr int32) {
if code.Len() == 0 || code.Len() > functionAlignment {
panic("inconsistency")
}
code.Align(functionAlignment, paddingByte)
Add.opImm(code, types.I64, regStackLimit, gen.StackReserve)
var notResume links.L
Test.opFromReg(code, types.I64, regResult, regResult)
Je.rel8.opStub(code)
notResume.AddSite(code.Len())
Ret.op(code) // simulate return from snapshot function call
notResume.Addr = code.Len()
mach.updateBranches8(code, ¬Resume)
CallRel.op(code, startAddr)
return code.Len()
}
func (mach X86) OpBranchIf(code gen.Coder, x values.Operand, yes bool, addr int32) (sites []int32) {
var cond values.Condition
if x.Storage == values.ConditionFlags {
cond = x.Condition()
} else {
reg, _, own := mach.opBorrowMaybeScratchReg(code, x, false)
if own {
defer code.FreeReg(types.I32, reg)
}
Test.opFromReg(code, types.I32, reg, reg)
cond = values.Ne
}
if !yes {
cond = values.InvertedConditions[cond]
}
var end links.L
switch {
case cond >= values.MinUnorderedOrCondition:
Jp.op(code, addr)
sites = append(sites, code.Len())
case cond >= values.MinOrderedAndCondition:
Jp.rel8.opStub(code)
end.AddSite(code.Len())
}
conditionInsns[cond].jcc.op(code, addr)
sites = append(sites, code.Len())
end.Addr = code.Len()
mach.updateBranches8(code, &end)
return
}
// UpdateCalls modifies CallRel instructions, possibly while they are being
// executed.
func (mach X86) UpdateCalls(code gen.OpCoder, l *links.L) {
funcAddr := l.FinalAddr()
for _, retAddr := range l.Sites {
mach.PutUint32(code.Bytes()[retAddr-4:retAddr], uint32(funcAddr-retAddr))
}
}
// updateBranches8 modifies 8-bit relocations of Jmp and Jcc instructions.
func (mach X86) updateBranches8(code gen.OpCoder, l *links.L) {
labelAddr := l.FinalAddr()
for _, retAddr := range l.Sites {
mach.updateAddr8(code, retAddr, labelAddr-retAddr)
}
}
func (mach X86) OpSelect(code gen.RegCoder, a, b, condOperand values.Operand) values.Operand {
defer code.Consumed(condOperand)
var cond values.Condition
switch condOperand.Storage {
case values.VarMem:
Cmp.opImmToStack(code, types.I32, condOperand.VarMemOffset(), 0)
cond = values.Ne
case values.VarReg, values.TempReg:
reg := condOperand.Reg()
Test.opFromReg(code, types.I32, reg, reg)
cond = values.Ne
case values.Stack:
mach.OpAddImmToStackPtr(code, 8) // do before cmp to avoid overwriting flags
Cmp.opImmToStack(code, types.I32, -8, 0)
cond = values.Ne
case values.ConditionFlags:
cond = condOperand.Condition()
case values.Imm:
if condOperand.ImmValue() != 0 {
code.Consumed(b)
return a
} else {
code.Consumed(a)
return b
}
default:
panic(condOperand)
}
t := a.Type
targetReg, _ := mach.opMaybeResultReg(code, b, true)
switch t.Category() {
case types.Int:
cmov := conditionInsns[cond].cmov
switch a.Storage {
case values.VarMem:
cmov.opFromStack(code, t, targetReg, a.VarMemOffset())
default:
aReg, _, own := mach.opBorrowMaybeScratchReg(code, a, true)
if own {
defer code.FreeReg(t, aReg)
}
cmov.opFromReg(code, t, targetReg, aReg)
}
case types.Float:
var moveIt links.L
var end links.L
cond = values.InvertedConditions[cond]
notCondJump := conditionInsns[cond].jcc
switch {
case cond >= values.MinUnorderedOrCondition:
Jp.rel8.opStub(code) // move it if unordered
moveIt.AddSite(code.Len())
notCondJump.rel8.opStub(code) // break if not cond
end.AddSite(code.Len())
case cond >= values.MinOrderedAndCondition:
Jp.rel8.opStub(code) // break if unordered
end.AddSite(code.Len())
notCondJump.rel8.opStub(code) // break if not cond
end.AddSite(code.Len())
default:
notCondJump.rel8.opStub(code) // break if not cond
end.AddSite(code.Len())
}
moveIt.Addr = code.Len()
mach.updateBranches8(code, &moveIt)
mach.OpMove(code, targetReg, a, false)
end.Addr = code.Len()
mach.updateBranches8(code, &end)
default:
panic(t)
}
// cmov zero-extends the target unconditionally
return values.TempRegOperand(t, targetReg, true)
}
// OpMove must not update CPU's condition flags if preserveFlags is set.
//
// X86 implementation note: must not blindly rely on regScratch or regResult in
// this function because we may be moving to one of them.
func (mach X86) OpMove(code gen.Coder, targetReg regs.R, x values.Operand, preserveFlags bool) (zeroExt bool) {
switch x.Type.Category() {
case types.Int:
switch x.Storage {
case values.Imm:
if value := x.ImmValue(); value == 0 && !preserveFlags {
Xor.opFromReg(code, types.I32, targetReg, targetReg)
} else {
MovImm64.op(code, x.Type, targetReg, value)
}
zeroExt = true
case values.VarMem:
Mov.opFromStack(code, x.Type, targetReg, x.VarMemOffset())
zeroExt = true
case values.VarReg:
if sourceReg := x.Reg(); sourceReg != targetReg {
Mov.opFromReg(code, x.Type, targetReg, sourceReg)
zeroExt = true
}
case values.TempReg:
if sourceReg := x.Reg(); sourceReg != targetReg {
Mov.opFromReg(code, x.Type, targetReg, sourceReg)
zeroExt = true
} else if targetReg == regResult {
zeroExt = x.RegZeroExt()
} else {
panic("moving temporary integer register to itself")
}
case values.Stack:
Pop.op(code, targetReg)
case values.ConditionFlags:
if x.Type != types.I32 {
panic(x)
}
var end links.L
cond := x.Condition()
setcc := conditionInsns[cond].setcc
switch {
case cond >= values.MinUnorderedOrCondition:
MovImm.opImm(code, x.Type, targetReg, 1) // true
Jp.rel8.opStub(code) // if unordered, else
end.AddSite(code.Len()) //
setcc.opReg(code, targetReg) // cond
case cond >= values.MinOrderedAndCondition:
MovImm.opImm(code, x.Type, targetReg, 0) // false
Jp.rel8.opStub(code) // if unordered, else
end.AddSite(code.Len()) //
setcc.opReg(code, targetReg) // cond
default:
setcc.opReg(code, targetReg)
Movzx8.opFromReg(code, x.Type, targetReg, targetReg)
}
end.Addr = code.Len()
mach.updateBranches8(code, &end)
zeroExt = true
default:
panic(x)
}
case types.Float:
switch x.Storage {
case values.Imm:
if value := x.ImmValue(); value == 0 {
PxorSSE.opFromReg(code, x.Type, targetReg, targetReg)
} else {
MovImm64.op(code, x.Type, regScratch, value) // integer scratch register
MovSSE.opFromReg(code, x.Type, targetReg, regScratch)
}
case values.VarMem:
MovsSSE.opFromStack(code, x.Type, targetReg, x.VarMemOffset())
case values.VarReg:
if sourceReg := x.Reg(); sourceReg != targetReg {
MovsSSE.opFromReg(code, x.Type, targetReg, sourceReg)
}
case values.TempReg:
if sourceReg := x.Reg(); sourceReg != targetReg {
MovsSSE.opFromReg(code, x.Type, targetReg, sourceReg)
} else if targetReg != regResult {
panic("moving temporary float register to itself")
}
case values.Stack:
popFloatOp(code, x.Type, targetReg)
default:
panic(x)
}
default:
panic(x)
}
code.Consumed(x)
return
}
func (mach X86) binaryDivmulOp(code gen.RegCoder, index uint8, a, b values.Operand) values.Operand {
insn := binaryDivmulInsns[index]
t := a.Type
if b.Storage == values.Imm {
value := b.ImmValue()
switch {
case value == -1:
reg, _ := mach.opMaybeResultReg(code, a, false)
Neg.opReg(code, t, reg)
return values.TempRegOperand(t, reg, true)
case insn.shiftImm.defined() && value > 0 && isPowerOfTwo(uint64(value)):
reg, _ := mach.opMaybeResultReg(code, a, false)
insn.shiftImm.op(code, t, reg, log2(uint64(value)))
return values.TempRegOperand(t, reg, true)
}
}
division := (index & opers.DivmulMul) == 0
checkZero := true
checkOverflow := true
if b.Storage.IsReg() {
if b.Reg() == regResult {
newReg := regScratch
if division {
var ok bool
// can't use scratch reg as divisor since it contains the dividend high bits
newReg, ok = code.TryAllocReg(t)
if !ok {
// borrow a register which we don't need in this function
MovMMX.opFromReg(code, types.I64, regScratchMMX, regTextBase)
defer MovMMX.opToReg(code, types.I64, regTextBase, regScratchMMX)
newReg = regTextBase
}
}
Mov.opFromReg(code, t, newReg, regResult)
b = values.RegOperand(true, t, newReg)
}
} else {
if division && b.Storage == values.Imm {
value := b.ImmValue()
if value != 0 {
checkZero = false
}
if value != -1 {
checkOverflow = false
}
}
reg, ok := code.TryAllocReg(t)
if !ok {
// borrow a register which we don't need in this function
MovMMX.opFromReg(code, types.I64, regScratchMMX, regTextBase)
defer MovMMX.opToReg(code, types.I64, regTextBase, regScratchMMX)
reg = regTextBase
}
mach.OpMove(code, reg, b, true)
b = values.RegOperand(true, t, reg)
}
mach.OpMove(code, regResult, a, false)
remainder := (index & opers.DivmulRem) != 0
var doNot links.L
if division {
if checkZero {
mach.opCheckDivideByZero(code, t, b.Reg())
}
if a.Storage == values.Imm {
value := a.ImmValue()
if t.Size() == types.Size32 {
if value != -0x80000000 {
checkOverflow = false
}
} else {
if value != -0x8000000000000000 {
checkOverflow = false
}
}
}
signed := (index & opers.DivmulSign) != 0
if signed && checkOverflow {
var do links.L
if remainder {
Xor.opFromReg(code, types.I32, regScratch, regScratch) // moved to result at the end
Cmp.opImm(code, t, b.Reg(), -1)
Je.rel8.opStub(code)
doNot.AddSite(code.Len())
} else {
switch t.Size() {
case types.Size32:
Cmp.opImm(code, t, regResult, -0x80000000)
case types.Size64:
MovImm64.op(code, t, regScratch, -0x8000000000000000)
Cmp.opFromReg(code, t, regResult, regScratch)
default:
panic(a)
}
Jne.rel8.opStub(code)
do.AddSite(code.Len())
Cmp.opImm(code, t, b.Reg(), -1)
Jne.rel8.opStub(code)
do.AddSite(code.Len())
code.OpTrapCall(traps.IntegerOverflow)
}
do.Addr = code.Len()
mach.updateBranches8(code, &do)
}
if signed {
// sign-extend dividend low bits to high bits
CdqCqo.op(code, t)
} else {
// zero-extend dividend high bits
Xor.opFromReg(code, types.I32, regScratch, regScratch)
}
}
insn.opReg(code, t, b.Reg())
code.Consumed(b)
doNot.Addr = code.Len()
mach.updateBranches8(code, &doNot)
if remainder {
Mov.opFromReg(code, t, regResult, regScratch)
}
return values.TempRegOperand(t, regResult, true)
}
// StoreOp makes sure that index gets zero-extended if it's a VarReg operand.
func (mach X86) StoreOp(code gen.RegCoder, oper uint16, index, x values.Operand, offset uint32) {
size := oper >> 8
baseReg, indexReg, ownIndexReg, disp := mach.opMemoryAddress(code, size, index, offset)
if ownIndexReg {
defer code.FreeReg(types.I64, indexReg)
}
store := memoryStores[uint8(oper)]
insnType := store.insnType
if insnType == 0 {
insnType = x.Type
}
if x.Storage == values.Imm {
value := x.ImmValue()
value32 := int32(value)
switch {
case size == 1:
value32 = int32(int8(value32))
case size == 2:
value32 = int32(int16(value32))
case size == 4 || (value >= -0x80000000 && value < 0x80000000):
default:
goto large
}
store.insn.opImmToIndirect(code, insnType, 0, indexReg, baseReg, disp, value32)
return
large:
}
valueReg, _, own := mach.opBorrowMaybeResultReg(code, x, false)
if own {
defer code.FreeReg(x.Type, valueReg)
}
store.insn.opToIndirect(code, insnType, valueReg, 0, indexReg, baseReg, disp)
}
// opMemoryAddress may return the scratch register as the base.
func (mach X86) opMemoryAddress(code gen.RegCoder, size uint16, index values.Operand, offset uint32) (baseReg, indexReg regs.R, ownIndexReg bool, disp int32) {
sizeReach := uint64(size - 1)
reachOffset := uint64(offset) + sizeReach
if reachOffset >= 0x80000000 {
code.OpTrapCall(traps.MemoryOutOfBounds)
return
}
alreadyChecked := reachOffset < uint64(index.Bounds.Upper)
switch index.Storage {
case values.Imm:
value := uint64(index.ImmValue())
if value >= 0x80000000 {
code.OpTrapCall(traps.MemoryOutOfBounds)
return
}
addr := value + uint64(offset)
reachAddr := addr + sizeReach
if reachAddr >= 0x80000000 {
code.OpTrapCall(traps.MemoryOutOfBounds)
return
}
if reachAddr < uint64(code.MinMemorySize()) || alreadyChecked {
baseReg = regMemoryBase
indexReg = NoIndex
disp = int32(addr)
return
}
Lea.opFromIndirect(code, types.I64, regScratch, 0, NoIndex, regMemoryBase, int32(reachAddr))
default:
reg, zeroExt, own := mach.opBorrowMaybeScratchReg(code, index, true)
if !zeroExt {
Mov.opFromReg(code, types.I32, reg, reg) // zero-extend index
}
if alreadyChecked {
baseReg = regMemoryBase
indexReg = reg
ownIndexReg = own
disp = int32(offset)
return
}
Lea.opFromIndirect(code, types.I64, regScratch, 0, reg, regMemoryBase, int32(reachOffset))
if own {
code.FreeReg(types.I32, reg)
}
}
Cmp.opFromReg(code, types.I64, regScratch, regMemoryLimit)
if addr := code.TrapTrampolineAddr(traps.MemoryOutOfBounds); addr != 0 {
Jge.op(code, addr)
} else {
var checked links.L
Jl.rel8.opStub(code)
checked.AddSite(code.Len())
code.OpTrapCall(traps.MemoryOutOfBounds)
checked.Addr = code.Len()
mach.updateBranches8(code, &checked)
}
baseReg = regScratch
indexReg = NoIndex
disp = -int32(sizeReach)
return
}
func (mach X86) OpCurrentMemory(code gen.RegCoder) values.Operand {
Mov.opFromReg(code, types.I64, regResult, regMemoryLimit)
Sub.opFromReg(code, types.I64, regResult, regMemoryBase)
ShrImm.op(code, types.I64, regResult, wasm.PageBits)
return values.TempRegOperand(types.I32, regResult, true)
}
func (mach X86) OpGrowMemory(code gen.RegCoder, x values.Operand) values.Operand {
var out links.L
var fail links.L
MovMMX.opToReg(code, types.I64, regScratch, regMemoryGrowLimitMMX)
targetReg, zeroExt := mach.opMaybeResultReg(code, x, false)
if !zeroExt {
Mov.opFromReg(code, types.I32, targetReg, targetReg)
}
ShlImm.op(code, types.I64, targetReg, wasm.PageBits)
Add.opFromReg(code, types.I64, targetReg, regMemoryLimit) // new memory limit
Cmp.opFromReg(code, types.I64, targetReg, regScratch)
Jg.rel8.opStub(code)
fail.AddSite(code.Len())
Mov.opFromReg(code, types.I64, regScratch, regMemoryLimit)
Mov.opFromReg(code, types.I64, regMemoryLimit, targetReg)
Sub.opFromReg(code, types.I64, regScratch, regMemoryBase)
ShrImm.op(code, types.I64, regScratch, wasm.PageBits) // value on success
Mov.opFromReg(code, types.I32, targetReg, regScratch)
JmpRel.rel8.opStub(code)
out.AddSite(code.Len())
fail.Addr = code.Len()
mach.updateBranches8(code, &fail)
MovImm.opImm(code, types.I32, targetReg, -1) // value on failure
out.Addr = code.Len()
mach.updateBranches8(code, &out)
return values.TempRegOperand(types.I32, targetReg, true)
}