TrueCloudLab/neoneo-go
7
0
Fork 2
Code Issues Pull requests Projects Releases Packages Wiki Activity

compiler: emit short jumps where possible

Browse source 
Convert long jumps to short ones if the offset
can be represented in a single byte.

Close #805.
This commit is contained in:
Evgenii Stratonikov 2020-08-19 12:36:17 +03:00
parent 681e81420a
commit cfa62e7051
2 changed files with 246 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

135
pkg/compiler/codegen.go
View file

@ -1613,8 +1613,8 @@ func CodeGen(info *buildInfo) ([]byte, *DebugInfo, error) {
return nil, nil, err
}
buf := c.prog.Bytes()
if err := c.writeJumps(buf); err != nil {
buf, err := c.writeJumps(c.prog.Bytes())
if err != nil {
return nil, nil, err
}
return buf, c.emitDebugInfo(buf), nil
@ -1629,15 +1629,14 @@ func (c *codegen) resolveFuncDecls(f *ast.File, pkg *types.Package) {
}
}
func (c *codegen) writeJumps(b []byte) error {
func (c *codegen) writeJumps(b []byte) ([]byte, error) {
ctx := vm.NewContext(b)
var offsets []int
for op, _, err := ctx.Next(); err == nil && ctx.NextIP() < len(b); op, _, err = ctx.Next() {
switch op {
case opcode.JMP, opcode.JMPIFNOT, opcode.JMPIF, opcode.CALL,
opcode.JMPEQ, opcode.JMPNE,
opcode.JMPGT, opcode.JMPGE, opcode.JMPLE, opcode.JMPLT:
// Noop, assumed to be correct already. If you're fixing #905,
// make sure not to break "len" and "append" handling above.
case opcode.JMPL, opcode.JMPIFL, opcode.JMPIFNOTL,
opcode.JMPEQL, opcode.JMPNEL,
opcode.JMPGTL, opcode.JMPGEL, opcode.JMPLEL, opcode.JMPLTL,
@ -1648,15 +1647,135 @@ func (c *codegen) writeJumps(b []byte) error {
index := binary.LittleEndian.Uint16(arg)
if int(index) > len(c.l) {
return fmt.Errorf("unexpected label number: %d (max %d)", index, len(c.l))
return nil, fmt.Errorf("unexpected label number: %d (max %d)", index, len(c.l))
}
offset := c.l[index] - nextIP + 5
if offset > math.MaxInt32 || offset < math.MinInt32 {
return fmt.Errorf("label offset is too big at the instruction %d: %d (max %d, min %d)",
return nil, fmt.Errorf("label offset is too big at the instruction %d: %d (max %d, min %d)",
nextIP-5, offset, math.MaxInt32, math.MinInt32)
}
if op != opcode.PUSHA && math.MinInt8 <= offset && offset <= math.MaxInt8 {
offsets = append(offsets, ctx.IP())
}
binary.LittleEndian.PutUint32(arg, uint32(offset))
}
}
return nil
// Correct function ip range.
// Note: indices are sorted in increasing order.
for _, f := range c.funcs {
loop:
for _, ind := range offsets {
switch {
case ind > int(f.rng.End):
break loop
case ind < int(f.rng.Start):
f.rng.Start -= longToShortRemoveCount
f.rng.End -= longToShortRemoveCount
case ind >= int(f.rng.Start):
f.rng.End -= longToShortRemoveCount
}
}
}
return shortenJumps(b, offsets), nil
}
// longToShortRemoveCount is a difference between short and long instruction sizes in bytes.
const longToShortRemoveCount = 3
// shortenJumps returns converts b to a program where all long JMP*/CALL* specified by absolute offsets,
// are replaced with their corresponding short counterparts. It panics if either b or offsets are invalid.
// This is done in 2 passes:
// 1. Alter jump offsets taking into account parts to be removed.
// 2. Perform actual removal of jump targets.
// Note: after jump offsets altering, there can appear new candidates for conversion.
// These are ignored for now.
func shortenJumps(b []byte, offsets []int) []byte {
if len(offsets) == 0 {
return b
}
// 1. Alter existing jump offsets.
ctx := vm.NewContext(b)
for op, _, err := ctx.Next(); err == nil && ctx.IP() < len(b); op, _, err = ctx.Next() {
// we can't use arg returned by ctx.Next() because it is copied
nextIP := ctx.NextIP()
ip := ctx.IP()
switch op {
case opcode.JMP, opcode.JMPIFNOT, opcode.JMPIF, opcode.CALL,
opcode.JMPEQ, opcode.JMPNE,
opcode.JMPGT, opcode.JMPGE, opcode.JMPLE, opcode.JMPLT:
offset := int(int8(b[nextIP-1]))
offset += calcOffsetCorrection(ip, ip+offset, offsets)
b[nextIP-1] = byte(offset)
case opcode.JMPL, opcode.JMPIFL, opcode.JMPIFNOTL,
opcode.JMPEQL, opcode.JMPNEL,
opcode.JMPGTL, opcode.JMPGEL, opcode.JMPLEL, opcode.JMPLTL,
opcode.CALLL, opcode.PUSHA:
arg := b[nextIP-4:]
offset := int(int32(binary.LittleEndian.Uint32(arg)))
offset += calcOffsetCorrection(ip, ip+offset, offsets)
binary.LittleEndian.PutUint32(arg, uint32(offset))
}
}
// 2. Convert instructions.
copyOffset := 0
l := len(offsets)
b[offsets[0]] = toShortForm(b[offsets[0]])
for i := 0; i < l; i++ {
start := offsets[i] + 2
end := len(b)
if i != l-1 {
end = offsets[i+1] + 2
b[offsets[i+1]] = toShortForm(b[offsets[i+1]])
}
copy(b[start-copyOffset:], b[start+3:end])
copyOffset += longToShortRemoveCount
}
return b[:len(b)-copyOffset]
}
func calcOffsetCorrection(ip, target int, offsets []int) int {
cnt := 0
start := sort.Search(len(offsets), func(i int) bool {
return offsets[i] >= ip || offsets[i] >= target
})
for i := start; i < len(offsets) && (offsets[i] < target || offsets[i] <= ip); i++ {
ind := offsets[i]
if ip <= ind && ind < target ||
ind != ip && target <= ind && ind <= ip {
cnt += longToShortRemoveCount
}
}
if ip < target {
return -cnt
}
return cnt
}
func toShortForm(b byte) byte {
switch op := opcode.Opcode(b); op {
case opcode.JMPL:
return byte(opcode.JMP)
case opcode.JMPIFL:
return byte(opcode.JMPIF)
case opcode.JMPIFNOTL:
return byte(opcode.JMPIFNOT)
case opcode.JMPEQL:
return byte(opcode.JMPEQ)
case opcode.JMPNEL:
return byte(opcode.JMPNE)
case opcode.JMPGTL:
return byte(opcode.JMPGT)
case opcode.JMPGEL:
return byte(opcode.JMPGE)
case opcode.JMPLEL:
return byte(opcode.JMPLE)
case opcode.JMPLTL:
return byte(opcode.JMPLT)
case opcode.CALLL:
return byte(opcode.CALL)
default:
panic(fmt.Errorf("invalid opcode: %s", op))
}
}