cycle collector (#14071)

* figured out the wrong cycle trace proc problem
* cycle collector/break refactorings and minor improvements
This commit is contained in:
Andreas Rumpf
2020-04-22 17:34:35 +02:00
committed by GitHub
parent 01523b2b58
commit 269a458d74
6 changed files with 262 additions and 110 deletions

View File

@@ -1297,7 +1297,8 @@ proc genHook(m: BModule; t: PType; info: TLineInfo; op: TTypeAttachedOp): Rope =
if op == attachedTrace and m.config.selectedGC == gcOrc and
containsGarbageCollectedRef(t):
when false:
# re-enable this check
# unfortunately this check is wrong for an object type that only contains
# .cursor fields like 'Node' inside 'cycleleak'.
internalError(m.config, info, "no attached trace proc found")
result = rope("NIM_NIL")

View File

@@ -91,7 +91,7 @@ proc genWhileLoop(c: var TLiftCtx; i, dest: PNode): PNode =
proc genIf(c: var TLiftCtx; cond, action: PNode): PNode =
result = newTree(nkIfStmt, newTree(nkElifBranch, cond, action))
proc genContainerOf(c: TLiftCtx; objType: PType, field, x: PSym): PNode =
proc genContainerOf(c: TLiftCtx; objType: PType, field, x: PSym): PNode =
# generate: cast[ptr ObjType](cast[int](addr(x)) - offsetOf(objType.field))
let intType = getSysType(c.g, unknownLineInfo, tyInt)
@@ -104,7 +104,7 @@ proc genContainerOf(c: TLiftCtx; objType: PType, field, x: PSym): PNode =
let dotExpr = newNodeIT(nkDotExpr, c.info, x.typ)
dotExpr.add newNodeIT(nkType, c.info, objType)
dotExpr.add newSymNode(field)
let offsetOf = genBuiltin(c.g, mOffsetOf, "offsetof", dotExpr)
offsetOf.typ = intType
@@ -174,7 +174,7 @@ proc fillBodyObj(c: var TLiftCtx; n, body, x, y: PNode; enforceDefaultOp: bool)
caseStmt.add(branch)
if emptyBranches != n.len-1:
body.add(caseStmt)
c.filterDiscriminator = oldfilterDiscriminator
c.filterDiscriminator = oldfilterDiscriminator
of nkRecList:
for t in items(n): fillBodyObj(c, t, body, x, y, enforceDefaultOp)
else:

View File

@@ -0,0 +1,55 @@
#
#
# Nim's Runtime Library
# (c) Copyright 2019 Andreas Rumpf
#
# See the file "copying.txt", included in this
# distribution, for details about the copyright.
#
# Cell seqs for cyclebreaker and cyclicrefs_v2.
type
CellTuple = (ptr pointer, PNimType)
CellArray = ptr UncheckedArray[CellTuple]
CellSeq = object
len, cap: int
d: CellArray
proc add(s: var CellSeq, c: ptr pointer; t: PNimType) {.inline.} =
if s.len >= s.cap:
s.cap = s.cap * 3 div 2
when defined(useMalloc):
var d = cast[CellArray](c_malloc(uint(s.cap * sizeof(CellTuple))))
else:
var d = cast[CellArray](alloc(s.cap * sizeof(CellTuple)))
copyMem(d, s.d, s.len * sizeof(CellTuple))
when defined(useMalloc):
c_free(s.d)
else:
dealloc(s.d)
s.d = d
# XXX: realloc?
s.d[s.len] = (c, t)
inc(s.len)
proc init(s: var CellSeq, cap: int = 1024) =
s.len = 0
s.cap = cap
when defined(useMalloc):
s.d = cast[CellArray](c_malloc(uint(s.cap * sizeof(CellTuple))))
else:
s.d = cast[CellArray](alloc(s.cap * sizeof(CellTuple)))
proc deinit(s: var CellSeq) =
when defined(useMalloc):
c_free(s.d)
else:
dealloc(s.d)
s.d = nil
s.len = 0
s.cap = 0
proc pop(s: var CellSeq): (ptr pointer, PNimType) =
result = s.d[s.len-1]
dec s.len

View File

@@ -52,6 +52,8 @@ That seems acceptable, no leak is produced. This implies that the standard
depth-first traversal suffices.
]#
include cellseqs_v2
const
colGreen = 0b000
colYellow = 0b001
@@ -72,57 +74,9 @@ proc nimIncRefCyclic(p: pointer) {.compilerRtl, inl.} =
inc h.rc, rcIncrement
type
CellTuple = (ptr pointer, PNimType)
CellArray = ptr UncheckedArray[CellTuple]
CellSeq = object
len, cap: int
d: CellArray
GcEnv = object
traceStack: CellSeq
# ------------------- cell seq handling --------------------------------------
proc add(s: var CellSeq, c: ptr pointer; t: PNimType) {.inline.} =
if s.len >= s.cap:
s.cap = s.cap * 3 div 2
when defined(useMalloc):
var d = cast[CellArray](c_malloc(uint(s.cap * sizeof(CellTuple))))
else:
var d = cast[CellArray](alloc(s.cap * sizeof(CellTuple)))
copyMem(d, s.d, s.len * sizeof(CellTuple))
when defined(useMalloc):
c_free(s.d)
else:
dealloc(s.d)
s.d = d
# XXX: realloc?
s.d[s.len] = (c, t)
inc(s.len)
proc init(s: var CellSeq, cap: int = 1024) =
s.len = 0
s.cap = cap
when defined(useMalloc):
s.d = cast[CellArray](c_malloc(uint(s.cap * sizeof(CellTuple))))
else:
s.d = cast[CellArray](alloc(s.cap * sizeof(CellTuple)))
proc deinit(s: var CellSeq) =
when defined(useMalloc):
c_free(s.d)
else:
dealloc(s.d)
s.d = nil
s.len = 0
s.cap = 0
proc pop(s: var CellSeq): (ptr pointer, PNimType) =
result = s.d[s.len-1]
dec s.len
# ----------------------------------------------------------------------------
proc trace(p: pointer; desc: PNimType; j: var GcEnv) {.inline.} =
when false:
cprintf("[Trace] desc: %p %p\n", desc, p)

View File

@@ -14,6 +14,8 @@
# "Cyclic reference counting" by Rafael Dueire Lins
# R.D. Lins / Information Processing Letters 109 (2008) 7178
include cellseqs_v2
const
colGreen = 0b000
colYellow = 0b001
@@ -45,58 +47,10 @@ proc markCyclic*[T](x: ref T) {.inline.} =
h.setColor colYellow
type
CellTuple = (Cell, PNimType)
CellArray = ptr UncheckedArray[CellTuple]
CellSeq = object
len, cap: int
d: CellArray
GcEnv = object
traceStack: CellSeq
jumpStack: CellSeq
# ------------------- cell seq handling --------------------------------------
proc add(s: var CellSeq, c: Cell; t: PNimType) {.inline.} =
if s.len >= s.cap:
s.cap = s.cap * 3 div 2
when defined(useMalloc):
var d = cast[CellArray](c_malloc(uint(s.cap * sizeof(CellTuple))))
else:
var d = cast[CellArray](alloc(s.cap * sizeof(CellTuple)))
copyMem(d, s.d, s.len * sizeof(CellTuple))
when defined(useMalloc):
c_free(s.d)
else:
dealloc(s.d)
s.d = d
# XXX: realloc?
s.d[s.len] = (c, t)
inc(s.len)
proc init(s: var CellSeq, cap: int = 1024) =
s.len = 0
s.cap = cap
when defined(useMalloc):
s.d = cast[CellArray](c_malloc(uint(s.cap * sizeof(CellTuple))))
else:
s.d = cast[CellArray](alloc(s.cap * sizeof(CellTuple)))
proc deinit(s: var CellSeq) =
when defined(useMalloc):
c_free(s.d)
else:
dealloc(s.d)
s.d = nil
s.len = 0
s.cap = 0
proc pop(s: var CellSeq): (Cell, PNimType) =
result = s.d[s.len-1]
dec s.len
# ----------------------------------------------------------------------------
proc trace(s: Cell; desc: PNimType; j: var GcEnv) {.inline.} =
if desc.traceImpl != nil:
var p = s +! sizeof(RefHeader)
@@ -116,7 +70,10 @@ proc collect(s: Cell; desc: PNimType; j: var GcEnv) =
s.setColor colGreen
trace(s, desc, j)
while j.traceStack.len > 0:
let (t, desc) = j.traceStack.pop()
let (p, desc) = j.traceStack.pop()
let t = head(p[])
#Remove(<S, T>):
p[] = nil
if t.color == colRed:
t.setColor colGreen
trace(t, desc, j)
@@ -129,7 +86,8 @@ proc markRed(s: Cell; desc: PNimType; j: var GcEnv) =
s.setColor colRed
trace(s, desc, j)
while j.traceStack.len > 0:
let (t, desc) = j.traceStack.pop()
let (p, desc) = j.traceStack.pop()
let t = head(p[])
when traceCollector:
cprintf("[Cycle dec] %p %ld color %ld in jumpstack %ld\n", t, t.rc shr rcShift, t.color, t.rc and jumpStackFlag)
dec t.rc, rcIncrement
@@ -137,7 +95,7 @@ proc markRed(s: Cell; desc: PNimType; j: var GcEnv) =
t.rc = t.rc or jumpStackFlag
when traceCollector:
cprintf("[Now in jumpstack] %p %ld color %ld in jumpstack %ld\n", t, t.rc shr rcShift, t.color, t.rc and jumpStackFlag)
j.jumpStack.add(t, desc)
j.jumpStack.add(p, desc)
if t.color != colRed:
t.setColor colRed
trace(t, desc, j)
@@ -146,7 +104,8 @@ proc scanGreen(s: Cell; desc: PNimType; j: var GcEnv) =
s.setColor colGreen
trace(s, desc, j)
while j.traceStack.len > 0:
let (t, desc) = j.traceStack.pop()
let (p, desc) = j.traceStack.pop()
let t = head(p[])
if t.color != colGreen:
t.setColor colGreen
trace(t, desc, j)
@@ -158,15 +117,13 @@ proc nimTraceRef(q: pointer; desc: PNimType; env: pointer) {.compilerRtl.} =
let p = cast[ptr pointer](q)
if p[] != nil:
var j = cast[ptr GcEnv](env)
var t = head(p[])
j.traceStack.add(t, desc)
j.traceStack.add(p, desc)
proc nimTraceRefDyn(q: pointer; env: pointer) {.compilerRtl.} =
let p = cast[ptr pointer](q)
if p[] != nil:
var j = cast[ptr GcEnv](env)
var t = head(p[])
j.traceStack.add(t, cast[ptr PNimType](p[])[])
j.traceStack.add(p, cast[ptr PNimType](p[])[])
proc scan(s: Cell; desc: PNimType; j: var GcEnv) =
when traceCollector:
@@ -181,7 +138,8 @@ proc scan(s: Cell; desc: PNimType; j: var GcEnv) =
# that are still alive; we also need to mark what they
# refer to as alive:
while j.jumpStack.len > 0:
let (t, desc) = j.jumpStack.pop
let (p, desc) = j.jumpStack.pop
let t = head(p[])
# not in jump stack anymore!
t.rc = t.rc and not jumpStackFlag
if t.color == colRed and (t.rc and not rcMask) >= 0:
@@ -202,7 +160,8 @@ proc traceCycle(s: Cell; desc: PNimType) {.noinline.} =
markRed(s, desc, j)
scan(s, desc, j)
while j.jumpStack.len > 0:
let (t, desc) = j.jumpStack.pop
let (p, desc) = j.jumpStack.pop
let t = head(p[])
# not in jump stack anymore!
t.rc = t.rc and not jumpStackFlag
deinit j.jumpStack

View File

@@ -0,0 +1,183 @@
discard """
cmd: "nim c --gc:orc -d:nimThinout -r $file"
output: '''The first 20 hammings are: 1 2 3 4 5 MEM IS 0'''
"""
# test Nim Hamming Number Lazy List algo with reference counts and not...
# compile with "-d:release -d:danger" and test with various
# memory managment GC's, allocators, threading, etc.
from times import epochTime
from math import log2
# implement our own basic BigInt so the bigints library isn't necessary...
type
BigInt = object
digits: seq[uint32]
let zeroBigInt = BigInt(digits: @[ 0'u32 ])
let oneBigInt = BigInt(digits: @[ 1'u32 ])
proc shladd(bi: var BigInt; n: int; a: BigInt) =
var cry = 0'u64
for i in 0 ..< min(bi.digits.len, a.digits.len):
cry += (bi.digits[i].uint64 shl n) + a.digits[i].uint64
bi.digits[i] = cry.uint32
cry = cry shr 32
if cry > 0'u64:
bi.digits.add cry.uint32
proc `$`(x: BigInt): string =
if x.digits.len == 0 or (x.digits.len == 1 and x.digits[0] == 0'u32):
return "0"
var n = x
var msd = n.digits.high
result = ""
while msd >= 0:
if n.digits[msd] == 0'u32: msd.dec; continue
var brw = 0.uint64
for i in countdown(msd, 0):
let dvdnd = n.digits[i].uint64 + (brw shl 32)
let q = dvdnd div 10'u64; brw = dvdnd - q*10'u64; n.digits[i] = q.uint32
result &= $brw
for i in 0 .. result.high shr 1:
let tmp = result[^(i + 1)]
result[^(i + 1)] = result[i]
result[i] = tmp
proc convertTrival2BigInt(tpl: (uint32, uint32, uint32)): BigInt =
result = oneBigInt
let (x2, x3, x5) = tpl
for _ in 1 .. x2: result.shladd 1, zeroBigInt
for _ in 1 .. x3: result.shladd 1, result
for _ in 1 .. x5: result.shladd 2, result
type LogRep = (float64, (uint32, uint32, uint32))
type LogRepf = proc(x: LogRep): LogRep
const one: LogRep = (0.0f64, (0u32, 0u32, 0u32))
proc `<`(me: LogRep, othr: LogRep): bool = me[0] < othr[0]
const lb2 = 1.0'f64
const lb3 = 3.0'f64.log2
const lb5 = 5.0'f64.log2
proc mul2(me: LogRep): LogRep =
let (lr, tpl) = me; let (x2, x3, x5) = tpl
(lr + lb2, (x2 + 1, x3, x5))
proc mul3(me: LogRep): LogRep =
let (lr, tpl) = me; let (x2, x3, x5) = tpl
(lr + lb3, (x2, x3 + 1, x5))
proc mul5(me: LogRep): LogRep =
let (lr, tpl) = me; let (x2, x3, x5) = tpl
(lr + lb5, (x2, x3, x5 + 1))
type
LazyList = ref object
hd: LogRep
tlf: proc(): LazyList {.closure.}
tl: LazyList
proc rest(ll: LazyList): LazyList = # not thread-safe; needs lock on thunk
if ll.tlf != nil:
ll.tl = ll.tlf()
ll.tlf = nil
ll.tl
iterator hamming(until: int): (uint32, uint32, uint32) =
proc merge(x, y: LazyList): LazyList =
let xh = x.hd
let yh = y.hd
if xh < yh: LazyList(hd: xh, tlf: proc(): auto = merge x.rest, y)
else: LazyList(hd: yh, tlf: proc(): auto = merge x, y.rest)
proc smult(mltf: LogRepf; s: LazyList): LazyList =
proc smults(ss: LazyList): LazyList =
LazyList(hd: ss.hd.mltf, tlf: proc(): auto = ss.rest.smults)
s.smults
proc unnsm(s: LazyList, mltf: LogRepf): LazyList =
var r: LazyList = nil
let frst = LazyList(hd: one, tlf: proc(): LazyList = r)
r = if s == nil: smult mltf, frst else: s.merge smult(mltf, frst)
r
var hmpll: LazyList = ((nil.unnsm mul5).unnsm mul3).unnsm mul2
# var hmpll: LazyList = nil; for m in [mul5, mul3, mul2]: echo one.m # ; hmpll = unnsm(hmpll, m)
yield one[1]
var cnt = 1
while hmpll != nil:
yield hmpll.hd[1]
hmpll = hmpll.rest # almost forever
cnt.inc
if cnt > until: break
#when declared(thinout):
thinout(hmpll)
proc main =
stdout.write "The first 20 hammings are: "
for h in hamming(4):
write stdout, h.convertTrival2BigInt, " "
for h in hamming(200):
discard h.convertTrival2BigInt
let mem = getOccupiedMem()
main()
echo "MEM IS ", getOccupiedMem() - mem
#[
result = (smults, :envP.:up)(rest(:envP.ss2))
proc anon =
var
:tmpD_284230
:tmpD_284233
:tmpD_284236
try:
`=sink_283407`(result_283502,
`=sink_283927`(:tmpD_284236, (smults_283495,
wasMoved_284234(:tmpD_284233)
`=_284014`(:tmpD_284233, :envP_283898.:up_283899)
:tmpD_284233))
:tmpD_284236(
`=sink_283407`(:tmpD_284230, rest_283366(:envP_283898.ss2_-283497))
:tmpD_284230))
finally:
`=destroy_283914`(:tmpD_284236)
`=destroy_283388`(:tmpD_284230)
proc smuls(ss: LazyList_283350; :envP_283891): LazyList_283350 =
var :env_283913
try:
`=destroy_283951`(:env_283913)
internalNew_43643(:env_283913)
`=_283401`(:env_283913.ss2_-283497, ss_283497)
:env_283913.:up_283899 = :envP_283891
`=sink_283407`(result_283498, LazyList_283350(hd_283353: :envP_283891.mltf1_-283492(
:env_283913.ss2_-283497.hd_283353), tlf_283356: (:anonymous_283499,
let blitTmp_284227 = :env_283913
wasMoved_284228(:env_283913)
blitTmp_284227)))
finally:
`=destroy_283951`(:env_283913)
proc smul =
var
:env_283969
:tmpD_284220
try:
`=destroy_284008`(:env_283969)
internalNew_43643(:env_283969)
`=_283976`(:env_283969.mltf1_-283492, mltf_283492)
proc smults(ss: LazyList_283350; :envP_283891): LazyList_283350 =
result_283498 = LazyList_283350(hd_283353: mltf_283492(ss_283497.hd_283353), tlf_283356: proc (
:envP_283898): auto_43100 = result_283502 = smults_283495(rest_283366(ss_283497)))
`=sink_283407`(result_283494,
`=sink_283927`(:tmpD_284220, (smults_283495,
let blitTmp_284218 = :env_283969
wasMoved_284219(:env_283969)
blitTmp_284218))
:tmpD_284220(s_283493))
finally:
`=destroy_283914`(:tmpD_284220)
`=destroy_284008`(:env_283969)
]#