fixes #3184
This commit is contained in:
parent
d691867c84
commit
dcf830bba9
4 changed files with 503 additions and 53 deletions
|
|
@ -27,15 +27,14 @@ const
|
||||||
|
|
||||||
type
|
type
|
||||||
PTrunk = ptr Trunk
|
PTrunk = ptr Trunk
|
||||||
Trunk {.final.} = object
|
Trunk = object
|
||||||
next: PTrunk # all nodes are connected with this pointer
|
next: PTrunk # all nodes are connected with this pointer
|
||||||
key: int # start address at bit 0
|
key: int # start address at bit 0
|
||||||
bits: array[0..IntsPerTrunk-1, int] # a bit vector
|
bits: array[0..IntsPerTrunk-1, int] # a bit vector
|
||||||
|
|
||||||
TrunkBuckets = array[0..255, PTrunk]
|
TrunkBuckets = array[0..255, PTrunk]
|
||||||
IntSet {.final.} = object
|
IntSet = object
|
||||||
data: TrunkBuckets
|
data: TrunkBuckets
|
||||||
{.deprecated: [TIntSet: IntSet, TTrunk: Trunk, TTrunkBuckets: TrunkBuckets].}
|
|
||||||
|
|
||||||
type
|
type
|
||||||
AlignType = BiggestFloat
|
AlignType = BiggestFloat
|
||||||
|
|
@ -64,8 +63,6 @@ type
|
||||||
next, prev: PBigChunk # chunks of the same (or bigger) size
|
next, prev: PBigChunk # chunks of the same (or bigger) size
|
||||||
align: int
|
align: int
|
||||||
data: AlignType # start of usable memory
|
data: AlignType # start of usable memory
|
||||||
{.deprecated: [TAlignType: AlignType, TFreeCell: FreeCell, TBaseChunk: BaseChunk,
|
|
||||||
TBigChunk: BigChunk, TSmallChunk: SmallChunk].}
|
|
||||||
|
|
||||||
template smallChunkOverhead(): expr = sizeof(SmallChunk)-sizeof(AlignType)
|
template smallChunkOverhead(): expr = sizeof(SmallChunk)-sizeof(AlignType)
|
||||||
template bigChunkOverhead(): expr = sizeof(BigChunk)-sizeof(AlignType)
|
template bigChunkOverhead(): expr = sizeof(BigChunk)-sizeof(AlignType)
|
||||||
|
|
@ -79,18 +76,18 @@ template bigChunkOverhead(): expr = sizeof(BigChunk)-sizeof(AlignType)
|
||||||
|
|
||||||
type
|
type
|
||||||
PLLChunk = ptr LLChunk
|
PLLChunk = ptr LLChunk
|
||||||
LLChunk {.pure.} = object ## *low-level* chunk
|
LLChunk = object ## *low-level* chunk
|
||||||
size: int # remaining size
|
size: int # remaining size
|
||||||
acc: int # accumulator
|
acc: int # accumulator
|
||||||
next: PLLChunk # next low-level chunk; only needed for dealloc
|
next: PLLChunk # next low-level chunk; only needed for dealloc
|
||||||
|
|
||||||
PAvlNode = ptr AvlNode
|
PAvlNode = ptr AvlNode
|
||||||
AvlNode {.pure, final.} = object
|
AvlNode = object
|
||||||
link: array[0..1, PAvlNode] # Left (0) and right (1) links
|
link: array[0..1, PAvlNode] # Left (0) and right (1) links
|
||||||
key, upperBound: int
|
key, upperBound: int
|
||||||
level: int
|
level: int
|
||||||
|
|
||||||
MemRegion {.final, pure.} = object
|
MemRegion = object
|
||||||
minLargeObj, maxLargeObj: int
|
minLargeObj, maxLargeObj: int
|
||||||
freeSmallChunks: array[0..SmallChunkSize div MemAlign-1, PSmallChunk]
|
freeSmallChunks: array[0..SmallChunkSize div MemAlign-1, PSmallChunk]
|
||||||
llmem: PLLChunk
|
llmem: PLLChunk
|
||||||
|
|
@ -99,6 +96,7 @@ type
|
||||||
freeChunksList: PBigChunk # XXX make this a datastructure with O(1) access
|
freeChunksList: PBigChunk # XXX make this a datastructure with O(1) access
|
||||||
chunkStarts: IntSet
|
chunkStarts: IntSet
|
||||||
root, deleted, last, freeAvlNodes: PAvlNode
|
root, deleted, last, freeAvlNodes: PAvlNode
|
||||||
|
locked: bool # if locked, we cannot free pages.
|
||||||
{.deprecated: [TLLChunk: LLChunk, TAvlNode: AvlNode, TMemRegion: MemRegion].}
|
{.deprecated: [TLLChunk: LLChunk, TAvlNode: AvlNode, TMemRegion: MemRegion].}
|
||||||
|
|
||||||
# shared:
|
# shared:
|
||||||
|
|
@ -234,7 +232,8 @@ proc isSmallChunk(c: PChunk): bool {.inline.} =
|
||||||
proc chunkUnused(c: PChunk): bool {.inline.} =
|
proc chunkUnused(c: PChunk): bool {.inline.} =
|
||||||
result = not c.used
|
result = not c.used
|
||||||
|
|
||||||
iterator allObjects(m: MemRegion): pointer {.inline.} =
|
iterator allObjects(m: var MemRegion): pointer {.inline.} =
|
||||||
|
m.locked = true
|
||||||
for s in elements(m.chunkStarts):
|
for s in elements(m.chunkStarts):
|
||||||
# we need to check here again as it could have been modified:
|
# we need to check here again as it could have been modified:
|
||||||
if s in m.chunkStarts:
|
if s in m.chunkStarts:
|
||||||
|
|
@ -252,6 +251,7 @@ iterator allObjects(m: MemRegion): pointer {.inline.} =
|
||||||
else:
|
else:
|
||||||
let c = cast[PBigChunk](c)
|
let c = cast[PBigChunk](c)
|
||||||
yield addr(c.data)
|
yield addr(c.data)
|
||||||
|
m.locked = false
|
||||||
|
|
||||||
proc iterToProc*(iter: typed, envType: typedesc; procName: untyped) {.
|
proc iterToProc*(iter: typed, envType: typedesc; procName: untyped) {.
|
||||||
magic: "Plugin", compileTime.}
|
magic: "Plugin", compileTime.}
|
||||||
|
|
@ -385,7 +385,7 @@ proc freeBigChunk(a: var MemRegion, c: PBigChunk) =
|
||||||
excl(a.chunkStarts, pageIndex(c))
|
excl(a.chunkStarts, pageIndex(c))
|
||||||
c = cast[PBigChunk](le)
|
c = cast[PBigChunk](le)
|
||||||
|
|
||||||
if c.size < ChunkOsReturn or doNotUnmap:
|
if c.size < ChunkOsReturn or doNotUnmap or a.locked:
|
||||||
incl(a, a.chunkStarts, pageIndex(c))
|
incl(a, a.chunkStarts, pageIndex(c))
|
||||||
updatePrevSize(a, c, c.size)
|
updatePrevSize(a, c, c.size)
|
||||||
listAdd(a.freeChunksList, c)
|
listAdd(a.freeChunksList, c)
|
||||||
|
|
@ -442,26 +442,29 @@ proc getSmallChunk(a: var MemRegion): PSmallChunk =
|
||||||
# -----------------------------------------------------------------------------
|
# -----------------------------------------------------------------------------
|
||||||
proc isAllocatedPtr(a: MemRegion, p: pointer): bool {.benign.}
|
proc isAllocatedPtr(a: MemRegion, p: pointer): bool {.benign.}
|
||||||
|
|
||||||
proc allocInv(a: MemRegion): bool =
|
when true:
|
||||||
## checks some (not all yet) invariants of the allocator's data structures.
|
template allocInv(a: MemRegion): bool = true
|
||||||
for s in low(a.freeSmallChunks)..high(a.freeSmallChunks):
|
else:
|
||||||
var c = a.freeSmallChunks[s]
|
proc allocInv(a: MemRegion): bool =
|
||||||
while not (c == nil):
|
## checks some (not all yet) invariants of the allocator's data structures.
|
||||||
if c.next == c:
|
for s in low(a.freeSmallChunks)..high(a.freeSmallChunks):
|
||||||
echo "[SYSASSERT] c.next == c"
|
var c = a.freeSmallChunks[s]
|
||||||
return false
|
while not (c == nil):
|
||||||
if not (c.size == s * MemAlign):
|
if c.next == c:
|
||||||
echo "[SYSASSERT] c.size != s * MemAlign"
|
echo "[SYSASSERT] c.next == c"
|
||||||
return false
|
|
||||||
var it = c.freeList
|
|
||||||
while not (it == nil):
|
|
||||||
if not (it.zeroField == 0):
|
|
||||||
echo "[SYSASSERT] it.zeroField != 0"
|
|
||||||
c_printf("%ld %p\n", it.zeroField, it)
|
|
||||||
return false
|
return false
|
||||||
it = it.next
|
if not (c.size == s * MemAlign):
|
||||||
c = c.next
|
echo "[SYSASSERT] c.size != s * MemAlign"
|
||||||
result = true
|
return false
|
||||||
|
var it = c.freeList
|
||||||
|
while not (it == nil):
|
||||||
|
if not (it.zeroField == 0):
|
||||||
|
echo "[SYSASSERT] it.zeroField != 0"
|
||||||
|
c_printf("%ld %p\n", it.zeroField, it)
|
||||||
|
return false
|
||||||
|
it = it.next
|
||||||
|
c = c.next
|
||||||
|
result = true
|
||||||
|
|
||||||
proc rawAlloc(a: var MemRegion, requestedSize: int): pointer =
|
proc rawAlloc(a: var MemRegion, requestedSize: int): pointer =
|
||||||
sysAssert(allocInv(a), "rawAlloc: begin")
|
sysAssert(allocInv(a), "rawAlloc: begin")
|
||||||
|
|
|
||||||
|
|
@ -190,6 +190,12 @@ proc prepareDealloc(cell: PCell) =
|
||||||
(cast[Finalizer](cell.typ.finalizer))(cellToUsr(cell))
|
(cast[Finalizer](cell.typ.finalizer))(cellToUsr(cell))
|
||||||
dec(gch.recGcLock)
|
dec(gch.recGcLock)
|
||||||
|
|
||||||
|
template beforeDealloc(gch: var GcHeap; c: PCell; msg: typed) =
|
||||||
|
when false:
|
||||||
|
for i in 0..gch.decStack.len-1:
|
||||||
|
if gch.decStack.d[i] == c:
|
||||||
|
sysAssert(false, msg)
|
||||||
|
|
||||||
proc rtlAddCycleRoot(c: PCell) {.rtl, inl.} =
|
proc rtlAddCycleRoot(c: PCell) {.rtl, inl.} =
|
||||||
# we MUST access gch as a global here, because this crosses DLL boundaries!
|
# we MUST access gch as a global here, because this crosses DLL boundaries!
|
||||||
when hasThreadSupport and hasSharedHeap:
|
when hasThreadSupport and hasSharedHeap:
|
||||||
|
|
@ -541,6 +547,7 @@ proc growObj(old: pointer, newsize: int, gch: var GcHeap): pointer =
|
||||||
d[j] = res
|
d[j] = res
|
||||||
break
|
break
|
||||||
dec(j)
|
dec(j)
|
||||||
|
beforeDealloc(gch, ol, "growObj stack trash")
|
||||||
rawDealloc(gch.region, ol)
|
rawDealloc(gch.region, ol)
|
||||||
else:
|
else:
|
||||||
# we split the old refcount in 2 parts. XXX This is still not entirely
|
# we split the old refcount in 2 parts. XXX This is still not entirely
|
||||||
|
|
@ -574,6 +581,7 @@ proc freeCyclicCell(gch: var GcHeap, c: PCell) =
|
||||||
when logGC: writeCell("cycle collector dealloc cell", c)
|
when logGC: writeCell("cycle collector dealloc cell", c)
|
||||||
when reallyDealloc:
|
when reallyDealloc:
|
||||||
sysAssert(allocInv(gch.region), "free cyclic cell")
|
sysAssert(allocInv(gch.region), "free cyclic cell")
|
||||||
|
beforeDealloc(gch, c, "freeCyclicCell: stack trash")
|
||||||
rawDealloc(gch.region, c)
|
rawDealloc(gch.region, c)
|
||||||
else:
|
else:
|
||||||
gcAssert(c.typ != nil, "freeCyclicCell")
|
gcAssert(c.typ != nil, "freeCyclicCell")
|
||||||
|
|
@ -601,16 +609,6 @@ when useMarkForDebug or useBackupGc:
|
||||||
proc markGlobals(gch: var GcHeap) =
|
proc markGlobals(gch: var GcHeap) =
|
||||||
for i in 0 .. < globalMarkersLen: globalMarkers[i]()
|
for i in 0 .. < globalMarkersLen: globalMarkers[i]()
|
||||||
|
|
||||||
proc stackMarkS(gch: var GcHeap, p: pointer) {.inline.} =
|
|
||||||
# the addresses are not as cells on the stack, so turn them to cells:
|
|
||||||
var cell = usrToCell(p)
|
|
||||||
var c = cast[ByteAddress](cell)
|
|
||||||
if c >% PageSize:
|
|
||||||
# fast check: does it look like a cell?
|
|
||||||
var objStart = cast[PCell](interiorAllocatedPtr(gch.region, cell))
|
|
||||||
if objStart != nil:
|
|
||||||
markS(gch, objStart)
|
|
||||||
|
|
||||||
when logGC:
|
when logGC:
|
||||||
var
|
var
|
||||||
cycleCheckA: array[100, PCell]
|
cycleCheckA: array[100, PCell]
|
||||||
|
|
@ -669,10 +667,6 @@ proc nimGCvisit(d: pointer, op: int) {.compilerRtl.} =
|
||||||
|
|
||||||
proc collectZCT(gch: var GcHeap): bool {.benign.}
|
proc collectZCT(gch: var GcHeap): bool {.benign.}
|
||||||
|
|
||||||
when useMarkForDebug or useBackupGc:
|
|
||||||
proc markStackAndRegistersForSweep(gch: var GcHeap) {.noinline, cdecl,
|
|
||||||
benign.}
|
|
||||||
|
|
||||||
proc collectCycles(gch: var GcHeap) =
|
proc collectCycles(gch: var GcHeap) =
|
||||||
when hasThreadSupport:
|
when hasThreadSupport:
|
||||||
for c in gch.toDispose:
|
for c in gch.toDispose:
|
||||||
|
|
@ -681,7 +675,10 @@ proc collectCycles(gch: var GcHeap) =
|
||||||
while gch.zct.len > 0: discard collectZCT(gch)
|
while gch.zct.len > 0: discard collectZCT(gch)
|
||||||
when useBackupGc:
|
when useBackupGc:
|
||||||
cellsetReset(gch.marked)
|
cellsetReset(gch.marked)
|
||||||
markStackAndRegistersForSweep(gch)
|
var d = gch.decStack.d
|
||||||
|
for i in 0..gch.decStack.len-1:
|
||||||
|
sysAssert isAllocatedPtr(gch.region, d[i]), "collectCycles"
|
||||||
|
markS(gch, d[i])
|
||||||
markGlobals(gch)
|
markGlobals(gch)
|
||||||
sweep(gch)
|
sweep(gch)
|
||||||
|
|
||||||
|
|
@ -710,10 +707,6 @@ include gc_common
|
||||||
proc markStackAndRegisters(gch: var GcHeap) {.noinline, cdecl.} =
|
proc markStackAndRegisters(gch: var GcHeap) {.noinline, cdecl.} =
|
||||||
forEachStackSlot(gch, gcMark)
|
forEachStackSlot(gch, gcMark)
|
||||||
|
|
||||||
when useMarkForDebug or useBackupGc:
|
|
||||||
proc markStackAndRegistersForSweep(gch: var GcHeap) =
|
|
||||||
forEachStackSlot(gch, stackMarkS)
|
|
||||||
|
|
||||||
proc collectZCT(gch: var GcHeap): bool =
|
proc collectZCT(gch: var GcHeap): bool =
|
||||||
# Note: Freeing may add child objects to the ZCT! So essentially we do
|
# Note: Freeing may add child objects to the ZCT! So essentially we do
|
||||||
# deep freeing, which is bad for incremental operation. In order to
|
# deep freeing, which is bad for incremental operation. In order to
|
||||||
|
|
@ -752,6 +745,7 @@ proc collectZCT(gch: var GcHeap): bool =
|
||||||
forAllChildren(c, waZctDecRef)
|
forAllChildren(c, waZctDecRef)
|
||||||
when reallyDealloc:
|
when reallyDealloc:
|
||||||
sysAssert(allocInv(gch.region), "collectZCT: rawDealloc")
|
sysAssert(allocInv(gch.region), "collectZCT: rawDealloc")
|
||||||
|
beforeDealloc(gch, c, "collectZCT: stack trash")
|
||||||
rawDealloc(gch.region, c)
|
rawDealloc(gch.region, c)
|
||||||
else:
|
else:
|
||||||
sysAssert(c.typ != nil, "collectZCT 2")
|
sysAssert(c.typ != nil, "collectZCT 2")
|
||||||
|
|
@ -811,11 +805,6 @@ proc collectCTBody(gch: var GcHeap) =
|
||||||
if gch.maxPause > 0 and duration > gch.maxPause:
|
if gch.maxPause > 0 and duration > gch.maxPause:
|
||||||
c_fprintf(c_stdout, "[GC] missed deadline: %ld\n", duration)
|
c_fprintf(c_stdout, "[GC] missed deadline: %ld\n", duration)
|
||||||
|
|
||||||
when useMarkForDebug or useBackupGc:
|
|
||||||
proc markForDebug(gch: var GcHeap) =
|
|
||||||
markStackAndRegistersForSweep(gch)
|
|
||||||
markGlobals(gch)
|
|
||||||
|
|
||||||
when defined(nimCoroutines):
|
when defined(nimCoroutines):
|
||||||
proc currentStackSizes(): int =
|
proc currentStackSizes(): int =
|
||||||
for stack in items(gch.stack):
|
for stack in items(gch.stack):
|
||||||
|
|
|
||||||
457
tests/gc/thavlak.nim
Normal file
457
tests/gc/thavlak.nim
Normal file
|
|
@ -0,0 +1,457 @@
|
||||||
|
discard """
|
||||||
|
output: '''Welcome to LoopTesterApp, Nim edition
|
||||||
|
Constructing Simple CFG...
|
||||||
|
15000 dummy loops
|
||||||
|
Constructing CFG...
|
||||||
|
Performing Loop Recognition
|
||||||
|
1 Iteration
|
||||||
|
Another 50 iterations...
|
||||||
|
..................................................
|
||||||
|
Found 1 loops (including artificial root node) (50)'''
|
||||||
|
"""
|
||||||
|
|
||||||
|
# bug #3184
|
||||||
|
|
||||||
|
import tables
|
||||||
|
import sequtils
|
||||||
|
import sets
|
||||||
|
|
||||||
|
type
|
||||||
|
BasicBlock = object
|
||||||
|
inEdges: seq[ref BasicBlock]
|
||||||
|
outEdges: seq[ref BasicBlock]
|
||||||
|
name: int
|
||||||
|
|
||||||
|
proc newBasicBlock(name: int): ref BasicBlock =
|
||||||
|
new(result)
|
||||||
|
result.inEdges = newSeq[ref BasicBlock]()
|
||||||
|
result.outEdges = newSeq[ref BasicBlock]()
|
||||||
|
result.name = name
|
||||||
|
|
||||||
|
proc hash(x: ref BasicBlock): int {.inline.} =
|
||||||
|
result = x.name
|
||||||
|
|
||||||
|
type
|
||||||
|
BasicBlockEdge = object
|
||||||
|
fr: ref BasicBlock
|
||||||
|
to: ref BasicBlock
|
||||||
|
|
||||||
|
Cfg = object
|
||||||
|
basicBlockMap: Table[int, ref BasicBlock]
|
||||||
|
edgeList: seq[BasicBlockEdge]
|
||||||
|
startNode: ref BasicBlock
|
||||||
|
|
||||||
|
proc newCfg(): Cfg =
|
||||||
|
result.basicBlockMap = initTable[int, ref BasicBlock]()
|
||||||
|
result.edgeList = newSeq[BasicBlockEdge]()
|
||||||
|
|
||||||
|
proc createNode(self: var Cfg, name: int): ref BasicBlock =
|
||||||
|
result = self.basicBlockMap.getOrDefault(name)
|
||||||
|
if result == nil:
|
||||||
|
result = newBasicBlock(name)
|
||||||
|
self.basicBlockMap.add name, result
|
||||||
|
|
||||||
|
if self.startNode == nil:
|
||||||
|
self.startNode = result
|
||||||
|
|
||||||
|
proc addEdge(self: var Cfg, edge: BasicBlockEdge) =
|
||||||
|
self.edgeList.add(edge)
|
||||||
|
|
||||||
|
proc getNumNodes(self: Cfg): int =
|
||||||
|
self.basicBlockMap.len
|
||||||
|
|
||||||
|
proc newBasicBlockEdge(cfg: var Cfg, fromName: int, toName: int): BasicBlockEdge =
|
||||||
|
result.fr = cfg.createNode(fromName)
|
||||||
|
result.to = cfg.createNode(toName)
|
||||||
|
result.fr.outEdges.add(result.to)
|
||||||
|
result.to.inEdges.add(result.fr)
|
||||||
|
cfg.addEdge(result)
|
||||||
|
|
||||||
|
type
|
||||||
|
SimpleLoop = object
|
||||||
|
basicBlocks: seq[ref BasicBlock] # TODO: set here
|
||||||
|
children: seq[ref SimpleLoop] # TODO: set here
|
||||||
|
parent: ref SimpleLoop
|
||||||
|
header: ref BasicBlock
|
||||||
|
isRoot: bool
|
||||||
|
isReducible: bool
|
||||||
|
counter: int
|
||||||
|
nestingLevel: int
|
||||||
|
depthLevel: int
|
||||||
|
|
||||||
|
proc newSimpleLoop(): ref SimpleLoop =
|
||||||
|
new(result)
|
||||||
|
result.basicBlocks = newSeq[ref BasicBlock]()
|
||||||
|
result.children = newSeq[ref SimpleLoop]()
|
||||||
|
result.parent = nil
|
||||||
|
result.header = nil
|
||||||
|
result.isRoot = false
|
||||||
|
result.isReducible = true
|
||||||
|
result.counter = 0
|
||||||
|
result.nestingLevel = 0
|
||||||
|
result.depthLevel = 0
|
||||||
|
|
||||||
|
proc addNode(self: ref SimpleLoop, bb: ref BasicBlock) =
|
||||||
|
self.basicBlocks.add bb
|
||||||
|
|
||||||
|
proc addChildLoop(self: ref SimpleLoop, loop: ref SimpleLoop) =
|
||||||
|
self.children.add loop
|
||||||
|
|
||||||
|
proc setParent(self: ref SimpleLoop, parent: ref SimpleLoop) =
|
||||||
|
self.parent = parent
|
||||||
|
self.parent.addChildLoop(self)
|
||||||
|
|
||||||
|
proc setHeader(self: ref SimpleLoop, bb: ref BasicBlock) =
|
||||||
|
self.basicBlocks.add(bb)
|
||||||
|
self.header = bb
|
||||||
|
|
||||||
|
proc setNestingLevel(self: ref SimpleLoop, level: int) =
|
||||||
|
self.nestingLevel = level
|
||||||
|
if level == 0: self.isRoot = true
|
||||||
|
|
||||||
|
var loop_counter: int = 0
|
||||||
|
|
||||||
|
type
|
||||||
|
Lsg = object
|
||||||
|
loops: seq[ref SimpleLoop]
|
||||||
|
root: ref SimpleLoop
|
||||||
|
|
||||||
|
proc createNewLoop(self: var Lsg): ref SimpleLoop =
|
||||||
|
result = newSimpleLoop()
|
||||||
|
loop_counter += 1
|
||||||
|
result.counter = loop_counter
|
||||||
|
|
||||||
|
proc addLoop(self: var Lsg, l: ref SimpleLoop) =
|
||||||
|
self.loops.add l
|
||||||
|
|
||||||
|
proc newLsg(): Lsg =
|
||||||
|
result.loops = newSeq[ref SimpleLoop]()
|
||||||
|
result.root = result.createNewLoop()
|
||||||
|
result.root.setNestingLevel(0)
|
||||||
|
result.addLoop(result.root)
|
||||||
|
|
||||||
|
proc getNumLoops(self: Lsg): int =
|
||||||
|
self.loops.len
|
||||||
|
|
||||||
|
type
|
||||||
|
UnionFindNode = object
|
||||||
|
parent: ref UnionFindNode
|
||||||
|
bb: ref BasicBlock
|
||||||
|
l: ref SimpleLoop
|
||||||
|
dfsNumber: int
|
||||||
|
|
||||||
|
proc newUnionFindNode(): ref UnionFindNode =
|
||||||
|
new(result)
|
||||||
|
when false:
|
||||||
|
result.parent = nil
|
||||||
|
result.bb = nil
|
||||||
|
result.l = nil
|
||||||
|
result.dfsNumber = 0
|
||||||
|
|
||||||
|
proc initNode(self: ref UnionFindNode, bb: ref BasicBlock, dfsNumber: int) =
|
||||||
|
self.parent = self
|
||||||
|
self.bb = bb
|
||||||
|
self.dfsNumber = dfsNumber
|
||||||
|
|
||||||
|
proc findSet(self: ref UnionFindNode): ref UnionFindNode =
|
||||||
|
var nodeList = newSeq[ref UnionFindNode]()
|
||||||
|
result = self
|
||||||
|
|
||||||
|
while result != result.parent:
|
||||||
|
var parent = result.parent
|
||||||
|
if parent != parent.parent: nodeList.add result
|
||||||
|
result = parent
|
||||||
|
|
||||||
|
for iter in nodeList: iter.parent = result.parent
|
||||||
|
|
||||||
|
proc union(self: ref UnionFindNode, unionFindNode: ref UnionFindNode) =
|
||||||
|
self.parent = unionFindNode
|
||||||
|
|
||||||
|
|
||||||
|
const
|
||||||
|
BB_TOP = 0 # uninitialized
|
||||||
|
BB_NONHEADER = 1 # a regular BB
|
||||||
|
BB_REDUCIBLE = 2 # reducible loop
|
||||||
|
BB_SELF = 3 # single BB loop
|
||||||
|
BB_IRREDUCIBLE = 4 # irreducible loop
|
||||||
|
BB_DEAD = 5 # a dead BB
|
||||||
|
BB_LAST = 6 # Sentinel
|
||||||
|
|
||||||
|
# # Marker for uninitialized nodes.
|
||||||
|
UNVISITED = -1
|
||||||
|
|
||||||
|
# # Safeguard against pathologic algorithm behavior.
|
||||||
|
MAXNONBACKPREDS = (32 * 1024)
|
||||||
|
|
||||||
|
type
|
||||||
|
HavlakLoopFinder = object
|
||||||
|
cfg: Cfg
|
||||||
|
lsg: Lsg
|
||||||
|
|
||||||
|
proc newHavlakLoopFinder(cfg: Cfg, lsg: Lsg): HavlakLoopFinder =
|
||||||
|
result.cfg = cfg
|
||||||
|
result.lsg = lsg
|
||||||
|
|
||||||
|
proc isAncestor(w: int, v: int, last: seq[int]): bool =
|
||||||
|
w <= v and v <= last[w]
|
||||||
|
|
||||||
|
proc dfs(currentNode: ref BasicBlock, nodes: var seq[ref UnionFindNode], number: var Table[ref BasicBlock, int], last: var seq[int], current: int): int =
|
||||||
|
var stack = @[(currentNode, current)]
|
||||||
|
while stack.len > 0:
|
||||||
|
let (currentNode, current) = stack.pop()
|
||||||
|
nodes[current].initNode(currentNode, current)
|
||||||
|
number[currentNode] = current
|
||||||
|
|
||||||
|
result = current
|
||||||
|
for target in currentNode.outEdges:
|
||||||
|
if number[target] == UNVISITED:
|
||||||
|
stack.add((target, result+1))
|
||||||
|
#result = dfs(target, nodes, number, last, result + 1)
|
||||||
|
last[number[currentNode]] = result
|
||||||
|
|
||||||
|
proc findLoops(self: var HavlakLoopFinder): int =
|
||||||
|
var startNode = self.cfg.startNode
|
||||||
|
if startNode == nil: return 0
|
||||||
|
var size = self.cfg.getNumNodes
|
||||||
|
|
||||||
|
var nonBackPreds = newSeq[HashSet[int]]()
|
||||||
|
var backPreds = newSeq[seq[int]]()
|
||||||
|
var number = initTable[ref BasicBlock, int]()
|
||||||
|
var header = newSeq[int](size)
|
||||||
|
var types = newSeq[int](size)
|
||||||
|
var last = newSeq[int](size)
|
||||||
|
var nodes = newSeq[ref UnionFindNode]()
|
||||||
|
|
||||||
|
for i in 1..size:
|
||||||
|
nonBackPreds.add initSet[int](1)
|
||||||
|
backPreds.add newSeq[int]()
|
||||||
|
nodes.add newUnionFindNode()
|
||||||
|
|
||||||
|
# Step a:
|
||||||
|
# - initialize all nodes as unvisited.
|
||||||
|
# - depth-first traversal and numbering.
|
||||||
|
# - unreached BB's are marked as dead.
|
||||||
|
#
|
||||||
|
for v in self.cfg.basicBlockMap.values: number[v] = UNVISITED
|
||||||
|
var res = dfs(startNode, nodes, number, last, 0)
|
||||||
|
|
||||||
|
# Step b:
|
||||||
|
# - iterate over all nodes.
|
||||||
|
#
|
||||||
|
# A backedge comes from a descendant in the DFS tree, and non-backedges
|
||||||
|
# from non-descendants (following Tarjan).
|
||||||
|
#
|
||||||
|
# - check incoming edges 'v' and add them to either
|
||||||
|
# - the list of backedges (backPreds) or
|
||||||
|
# - the list of non-backedges (nonBackPreds)
|
||||||
|
#
|
||||||
|
for w in 0 .. <size:
|
||||||
|
header[w] = 0
|
||||||
|
types[w] = BB_NONHEADER
|
||||||
|
|
||||||
|
var nodeW = nodes[w].bb
|
||||||
|
if nodeW != nil:
|
||||||
|
for nodeV in nodeW.inEdges:
|
||||||
|
var v = number[nodeV]
|
||||||
|
if v != UNVISITED:
|
||||||
|
if isAncestor(w, v, last):
|
||||||
|
backPreds[w].add v
|
||||||
|
else:
|
||||||
|
nonBackPreds[w].incl v
|
||||||
|
else:
|
||||||
|
types[w] = BB_DEAD
|
||||||
|
|
||||||
|
# Start node is root of all other loops.
|
||||||
|
header[0] = 0
|
||||||
|
|
||||||
|
# Step c:
|
||||||
|
#
|
||||||
|
# The outer loop, unchanged from Tarjan. It does nothing except
|
||||||
|
# for those nodes which are the destinations of backedges.
|
||||||
|
# For a header node w, we chase backward from the sources of the
|
||||||
|
# backedges adding nodes to the set P, representing the body of
|
||||||
|
# the loop headed by w.
|
||||||
|
#
|
||||||
|
# By running through the nodes in reverse of the DFST preorder,
|
||||||
|
# we ensure that inner loop headers will be processed before the
|
||||||
|
# headers for surrounding loops.
|
||||||
|
|
||||||
|
for w in countdown(size - 1, 0):
|
||||||
|
# this is 'P' in Havlak's paper
|
||||||
|
var nodePool = newSeq[ref UnionFindNode]()
|
||||||
|
|
||||||
|
var nodeW = nodes[w].bb
|
||||||
|
if nodeW != nil: # dead BB
|
||||||
|
# Step d:
|
||||||
|
for v in backPreds[w]:
|
||||||
|
if v != w:
|
||||||
|
nodePool.add nodes[v].findSet
|
||||||
|
else:
|
||||||
|
types[w] = BB_SELF
|
||||||
|
|
||||||
|
# Copy nodePool to workList.
|
||||||
|
#
|
||||||
|
var workList = newSeq[ref UnionFindNode]()
|
||||||
|
for x in nodePool: workList.add x
|
||||||
|
|
||||||
|
if nodePool.len != 0: types[w] = BB_REDUCIBLE
|
||||||
|
|
||||||
|
# work the list...
|
||||||
|
#
|
||||||
|
while workList.len > 0:
|
||||||
|
var x = workList[0]
|
||||||
|
workList.del(0)
|
||||||
|
|
||||||
|
# Step e:
|
||||||
|
#
|
||||||
|
# Step e represents the main difference from Tarjan's method.
|
||||||
|
# Chasing upwards from the sources of a node w's backedges. If
|
||||||
|
# there is a node y' that is not a descendant of w, w is marked
|
||||||
|
# the header of an irreducible loop, there is another entry
|
||||||
|
# into this loop that avoids w.
|
||||||
|
#
|
||||||
|
|
||||||
|
# The algorithm has degenerated. Break and
|
||||||
|
# return in this case.
|
||||||
|
#
|
||||||
|
var nonBackSize = nonBackPreds[x.dfsNumber].len
|
||||||
|
if nonBackSize > MAXNONBACKPREDS: return 0
|
||||||
|
|
||||||
|
for iter in nonBackPreds[x.dfsNumber]:
|
||||||
|
var y = nodes[iter]
|
||||||
|
var ydash = y.findSet
|
||||||
|
|
||||||
|
if not isAncestor(w, ydash.dfsNumber, last):
|
||||||
|
types[w] = BB_IRREDUCIBLE
|
||||||
|
nonBackPreds[w].incl ydash.dfsNumber
|
||||||
|
else:
|
||||||
|
if ydash.dfsNumber != w and not nodePool.contains(ydash):
|
||||||
|
workList.add ydash
|
||||||
|
nodePool.add ydash
|
||||||
|
|
||||||
|
# Collapse/Unionize nodes in a SCC to a single node
|
||||||
|
# For every SCC found, create a loop descriptor and link it in.
|
||||||
|
#
|
||||||
|
if (nodePool.len > 0) or (types[w] == BB_SELF):
|
||||||
|
var l = self.lsg.createNewLoop
|
||||||
|
|
||||||
|
l.setHeader(nodeW)
|
||||||
|
l.isReducible = types[w] != BB_IRREDUCIBLE
|
||||||
|
|
||||||
|
# At this point, one can set attributes to the loop, such as:
|
||||||
|
#
|
||||||
|
# the bottom node:
|
||||||
|
# iter = backPreds(w).begin();
|
||||||
|
# loop bottom is: nodes(iter).node;
|
||||||
|
#
|
||||||
|
# the number of backedges:
|
||||||
|
# backPreds(w).size()
|
||||||
|
#
|
||||||
|
# whether this loop is reducible:
|
||||||
|
# types(w) != BB_IRREDUCIBLE
|
||||||
|
#
|
||||||
|
nodes[w].l = l
|
||||||
|
|
||||||
|
for node in nodePool:
|
||||||
|
# Add nodes to loop descriptor.
|
||||||
|
header[node.dfsNumber] = w
|
||||||
|
node.union(nodes[w])
|
||||||
|
|
||||||
|
# Nested loops are not added, but linked together.
|
||||||
|
var node_l = node.l
|
||||||
|
if node_l != nil:
|
||||||
|
node_l.setParent(l)
|
||||||
|
else:
|
||||||
|
l.addNode(node.bb)
|
||||||
|
|
||||||
|
self.lsg.addLoop(l)
|
||||||
|
|
||||||
|
result = self.lsg.getNumLoops
|
||||||
|
|
||||||
|
|
||||||
|
type
|
||||||
|
LoopTesterApp = object
|
||||||
|
cfg: Cfg
|
||||||
|
lsg: Lsg
|
||||||
|
|
||||||
|
proc newLoopTesterApp(): LoopTesterApp =
|
||||||
|
result.cfg = newCfg()
|
||||||
|
result.lsg = newLsg()
|
||||||
|
|
||||||
|
proc buildDiamond(self: var LoopTesterApp, start: int): int =
|
||||||
|
var bb0 = start
|
||||||
|
var x1 = newBasicBlockEdge(self.cfg, bb0, bb0 + 1)
|
||||||
|
var x2 = newBasicBlockEdge(self.cfg, bb0, bb0 + 2)
|
||||||
|
var x3 = newBasicBlockEdge(self.cfg, bb0 + 1, bb0 + 3)
|
||||||
|
var x4 = newBasicBlockEdge(self.cfg, bb0 + 2, bb0 + 3)
|
||||||
|
result = bb0 + 3
|
||||||
|
|
||||||
|
proc buildConnect(self: var LoopTesterApp, start1: int, end1: int) =
|
||||||
|
var x1 = newBasicBlockEdge(self.cfg, start1, end1)
|
||||||
|
|
||||||
|
proc buildStraight(self: var LoopTesterApp, start: int, n: int): int =
|
||||||
|
for i in 0..n-1:
|
||||||
|
self.buildConnect(start + i, start + i + 1)
|
||||||
|
result = start + n
|
||||||
|
|
||||||
|
proc buildBaseLoop(self: var LoopTesterApp, from1: int): int =
|
||||||
|
var header = self.buildStraight(from1, 1)
|
||||||
|
var diamond1 = self.buildDiamond(header)
|
||||||
|
var d11 = self.buildStraight(diamond1, 1)
|
||||||
|
var diamond2 = self.buildDiamond(d11)
|
||||||
|
var footer = self.buildStraight(diamond2, 1)
|
||||||
|
|
||||||
|
self.buildConnect(diamond2, d11)
|
||||||
|
self.buildConnect(diamond1, header)
|
||||||
|
self.buildConnect(footer, from1)
|
||||||
|
result = self.buildStraight(footer, 1)
|
||||||
|
|
||||||
|
proc run(self: var LoopTesterApp) =
|
||||||
|
echo "Welcome to LoopTesterApp, Nim edition"
|
||||||
|
echo "Constructing Simple CFG..."
|
||||||
|
|
||||||
|
var x1 = self.cfg.createNode(0)
|
||||||
|
var x2 = self.buildBaseLoop(0)
|
||||||
|
var x3 = self.cfg.createNode(1)
|
||||||
|
self.buildConnect(0, 2)
|
||||||
|
|
||||||
|
echo "15000 dummy loops"
|
||||||
|
|
||||||
|
for i in 1..15000:
|
||||||
|
var h = newHavlakLoopFinder(self.cfg, newLsg())
|
||||||
|
var res = h.findLoops
|
||||||
|
|
||||||
|
echo "Constructing CFG..."
|
||||||
|
var n = 2
|
||||||
|
|
||||||
|
for parlooptrees in 1..10:
|
||||||
|
var x6 = self.cfg.createNode(n + 1)
|
||||||
|
self.buildConnect(2, n + 1)
|
||||||
|
n += 1
|
||||||
|
for i in 1..100:
|
||||||
|
var top = n
|
||||||
|
n = self.buildStraight(n, 1)
|
||||||
|
for j in 1..25: n = self.buildBaseLoop(n)
|
||||||
|
var bottom = self.buildStraight(n, 1)
|
||||||
|
self.buildConnect n, top
|
||||||
|
n = bottom
|
||||||
|
self.buildConnect(n, 1)
|
||||||
|
|
||||||
|
echo "Performing Loop Recognition\n1 Iteration"
|
||||||
|
|
||||||
|
var h = newHavlakLoopFinder(self.cfg, newLsg())
|
||||||
|
var loops = h.findLoops
|
||||||
|
|
||||||
|
echo "Another 50 iterations..."
|
||||||
|
|
||||||
|
var sum = 0
|
||||||
|
for i in 1..50:
|
||||||
|
write stdout, "."
|
||||||
|
flushFile(stdout)
|
||||||
|
var hlf = newHavlakLoopFinder(self.cfg, newLsg())
|
||||||
|
sum += hlf.findLoops
|
||||||
|
#echo getOccupiedMem()
|
||||||
|
echo "\nFound ", loops, " loops (including artificial root node) (", sum, ")"
|
||||||
|
|
||||||
|
var l = newLoopTesterApp()
|
||||||
|
l.run
|
||||||
|
|
@ -153,6 +153,7 @@ proc gcTests(r: var TResults, cat: Category, options: string) =
|
||||||
testWithoutMs "refarrayleak"
|
testWithoutMs "refarrayleak"
|
||||||
|
|
||||||
testWithoutBoehm "tlists"
|
testWithoutBoehm "tlists"
|
||||||
|
testWithoutBoehm "thavlak"
|
||||||
|
|
||||||
test "stackrefleak"
|
test "stackrefleak"
|
||||||
test "cyclecollector"
|
test "cyclecollector"
|
||||||
|
|
|
||||||
Loading…
Add table
Add a link
Reference in a new issue