2015-05-06 12:40:21 +00:00
|
|
|
// Copyright 2015 the V8 project authors. All rights reserved.
|
|
|
|
// Use of this source code is governed by a BSD-style license that can be
|
|
|
|
// found in the LICENSE file.
|
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
#include <set>
|
|
|
|
|
2019-05-22 07:55:37 +00:00
|
|
|
#include "src/execution/isolate.h"
|
2018-04-09 19:11:22 +00:00
|
|
|
#include "src/heap/factory-inl.h"
|
2018-12-17 17:01:48 +00:00
|
|
|
#include "src/objects/heap-number-inl.h"
|
2019-05-23 13:27:57 +00:00
|
|
|
#include "src/utils/identity-map.h"
|
2019-05-23 08:51:46 +00:00
|
|
|
#include "src/objects/objects.h"
|
2016-09-20 16:07:25 +00:00
|
|
|
#include "src/zone/zone.h"
|
2015-05-06 12:40:21 +00:00
|
|
|
#include "test/cctest/cctest.h"
|
|
|
|
|
|
|
|
namespace v8 {
|
|
|
|
namespace internal {
|
|
|
|
|
|
|
|
// Helper for testing. A "friend" of the IdentityMapBase class, it is able to
|
|
|
|
// "move" objects to simulate GC for testing the internals of the map.
|
|
|
|
class IdentityMapTester : public HandleAndZoneScope {
|
|
|
|
public:
|
2017-02-20 21:46:38 +00:00
|
|
|
IdentityMap<void*, ZoneAllocationPolicy> map;
|
2015-05-06 12:40:21 +00:00
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
IdentityMapTester() : map(heap(), ZoneAllocationPolicy(main_zone())) {}
|
2015-05-06 12:40:21 +00:00
|
|
|
|
|
|
|
Heap* heap() { return isolate()->heap(); }
|
|
|
|
Isolate* isolate() { return main_isolate(); }
|
|
|
|
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
void TestInsertFind(Handle<Object> key1, void* val1, Handle<Object> key2,
|
|
|
|
void* val2) {
|
2015-05-06 12:40:21 +00:00
|
|
|
CHECK_NULL(map.Find(key1));
|
|
|
|
CHECK_NULL(map.Find(key2));
|
|
|
|
|
|
|
|
// Set {key1} the first time.
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
auto find_result = map.FindOrInsert(key1);
|
|
|
|
CHECK_NOT_NULL(find_result.entry);
|
|
|
|
CHECK(!find_result.already_exists);
|
|
|
|
*find_result.entry = val1;
|
2015-05-06 12:40:21 +00:00
|
|
|
|
|
|
|
for (int i = 0; i < 3; i++) { // Get and find {key1} K times.
|
|
|
|
{
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
auto new_find_result = map.FindOrInsert(key1);
|
|
|
|
CHECK(new_find_result.already_exists);
|
|
|
|
CHECK_EQ(find_result.entry, new_find_result.entry);
|
|
|
|
CHECK_EQ(val1, *new_find_result.entry);
|
2015-05-06 12:40:21 +00:00
|
|
|
CHECK_NULL(map.Find(key2));
|
|
|
|
}
|
|
|
|
{
|
|
|
|
void** nentry = map.Find(key1);
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
CHECK_EQ(find_result.entry, nentry);
|
2015-05-06 12:40:21 +00:00
|
|
|
CHECK_EQ(val1, *nentry);
|
|
|
|
CHECK_NULL(map.Find(key2));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Set {key2} the first time.
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
auto find_result2 = map.FindOrInsert(key2);
|
|
|
|
CHECK_NOT_NULL(find_result2.entry);
|
|
|
|
CHECK(!find_result2.already_exists);
|
|
|
|
*find_result2.entry = val2;
|
2015-05-06 12:40:21 +00:00
|
|
|
|
|
|
|
for (int i = 0; i < 3; i++) { // Get and find {key1} and {key2} K times.
|
|
|
|
{
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
auto new_find_result = map.FindOrInsert(key2);
|
|
|
|
CHECK_EQ(find_result2.entry, new_find_result.entry);
|
|
|
|
CHECK_EQ(val2, *new_find_result.entry);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
{
|
|
|
|
void** nentry = map.Find(key2);
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
CHECK_EQ(find_result2.entry, nentry);
|
2015-05-06 12:40:21 +00:00
|
|
|
CHECK_EQ(val2, *nentry);
|
|
|
|
}
|
|
|
|
{
|
|
|
|
void** nentry = map.Find(key1);
|
|
|
|
CHECK_EQ(val1, *nentry);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
void TestFindDelete(Handle<Object> key1, void* val1, Handle<Object> key2,
|
|
|
|
void* val2) {
|
|
|
|
CHECK_NULL(map.Find(key1));
|
|
|
|
CHECK_NULL(map.Find(key2));
|
|
|
|
|
|
|
|
// Set {key1} and {key2} for the first time.
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
auto find_result1 = map.FindOrInsert(key1);
|
|
|
|
CHECK(!find_result1.already_exists);
|
|
|
|
CHECK_NOT_NULL(find_result1.entry);
|
|
|
|
*find_result1.entry = val1;
|
|
|
|
auto find_result2 = map.FindOrInsert(key2);
|
|
|
|
CHECK(!find_result1.already_exists);
|
|
|
|
CHECK_NOT_NULL(find_result2.entry);
|
|
|
|
*find_result2.entry = val2;
|
2017-02-20 21:46:38 +00:00
|
|
|
|
|
|
|
for (int i = 0; i < 3; i++) { // Find {key1} and {key2} 3 times.
|
|
|
|
{
|
|
|
|
void** nentry = map.Find(key2);
|
|
|
|
CHECK_EQ(val2, *nentry);
|
|
|
|
}
|
|
|
|
{
|
|
|
|
void** nentry = map.Find(key1);
|
|
|
|
CHECK_EQ(val1, *nentry);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Delete {key1}
|
2018-04-19 14:10:45 +00:00
|
|
|
void* deleted_entry_1;
|
|
|
|
CHECK(map.Delete(key1, &deleted_entry_1));
|
2017-02-20 21:46:38 +00:00
|
|
|
CHECK_NOT_NULL(deleted_entry_1);
|
|
|
|
deleted_entry_1 = val1;
|
|
|
|
|
|
|
|
for (int i = 0; i < 3; i++) { // Find {key1} and not {key2} 3 times.
|
|
|
|
{
|
|
|
|
void** nentry = map.Find(key1);
|
|
|
|
CHECK_NULL(nentry);
|
|
|
|
}
|
|
|
|
{
|
|
|
|
void** nentry = map.Find(key2);
|
|
|
|
CHECK_EQ(val2, *nentry);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Delete {key2}
|
2018-04-19 14:10:45 +00:00
|
|
|
void* deleted_entry_2;
|
|
|
|
CHECK(map.Delete(key2, &deleted_entry_2));
|
2017-02-20 21:46:38 +00:00
|
|
|
CHECK_NOT_NULL(deleted_entry_2);
|
|
|
|
deleted_entry_2 = val2;
|
|
|
|
|
|
|
|
for (int i = 0; i < 3; i++) { // Don't find {key1} and {key2} 3 times.
|
|
|
|
{
|
|
|
|
void** nentry = map.Find(key1);
|
|
|
|
CHECK_NULL(nentry);
|
|
|
|
}
|
|
|
|
{
|
|
|
|
void** nentry = map.Find(key2);
|
|
|
|
CHECK_NULL(nentry);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-05-06 12:40:21 +00:00
|
|
|
Handle<Smi> smi(int value) {
|
|
|
|
return Handle<Smi>(Smi::FromInt(value), isolate());
|
|
|
|
}
|
|
|
|
|
|
|
|
Handle<Object> num(double value) {
|
|
|
|
return isolate()->factory()->NewNumber(value);
|
|
|
|
}
|
|
|
|
|
|
|
|
void SimulateGCByIncrementingSmisBy(int shift) {
|
2017-02-20 21:46:38 +00:00
|
|
|
for (int i = 0; i < map.capacity_; i++) {
|
2018-10-18 23:44:29 +00:00
|
|
|
Address key = map.keys_[i];
|
|
|
|
if (!Internals::HasHeapObjectTag(key)) {
|
|
|
|
map.keys_[i] = Internals::IntToSmi(Internals::SmiValue(key) + shift);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
map.gc_counter_ = -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
void CheckFind(Handle<Object> key, void* value) {
|
|
|
|
void** entry = map.Find(key);
|
|
|
|
CHECK_NOT_NULL(entry);
|
|
|
|
CHECK_EQ(value, *entry);
|
|
|
|
}
|
|
|
|
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
void CheckFindOrInsert(Handle<Object> key, void* value) {
|
|
|
|
auto find_result = map.FindOrInsert(key);
|
|
|
|
CHECK(find_result.already_exists);
|
|
|
|
CHECK_NOT_NULL(find_result.entry);
|
|
|
|
CHECK_EQ(value, *find_result.entry);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
void CheckDelete(Handle<Object> key, void* value) {
|
2018-04-19 14:10:45 +00:00
|
|
|
void* entry;
|
|
|
|
CHECK(map.Delete(key, &entry));
|
2017-02-20 21:46:38 +00:00
|
|
|
CHECK_NOT_NULL(entry);
|
|
|
|
CHECK_EQ(value, entry);
|
|
|
|
}
|
|
|
|
|
2015-05-06 12:40:21 +00:00
|
|
|
void PrintMap() {
|
|
|
|
PrintF("{\n");
|
2017-02-20 21:46:38 +00:00
|
|
|
for (int i = 0; i < map.capacity_; i++) {
|
2015-05-06 12:40:21 +00:00
|
|
|
PrintF(" %3d: %p => %p\n", i, reinterpret_cast<void*>(map.keys_[i]),
|
|
|
|
reinterpret_cast<void*>(map.values_[i]));
|
|
|
|
}
|
|
|
|
PrintF("}\n");
|
|
|
|
}
|
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
void Resize() { map.Resize(map.capacity_ * 4); }
|
2015-05-06 12:40:21 +00:00
|
|
|
|
|
|
|
void Rehash() { map.Rehash(); }
|
|
|
|
};
|
|
|
|
|
|
|
|
TEST(Find_smi_not_found) {
|
|
|
|
IdentityMapTester t;
|
|
|
|
for (int i = 0; i < 100; i++) {
|
|
|
|
CHECK_NULL(t.map.Find(t.smi(i)));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
TEST(Find_num_not_found) {
|
|
|
|
IdentityMapTester t;
|
|
|
|
for (int i = 0; i < 100; i++) {
|
|
|
|
CHECK_NULL(t.map.Find(t.num(i + 0.2)));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
TEST(Delete_smi_not_found) {
|
|
|
|
IdentityMapTester t;
|
|
|
|
for (int i = 0; i < 100; i++) {
|
2018-04-19 14:10:45 +00:00
|
|
|
void* deleted_value = &t;
|
|
|
|
CHECK(!t.map.Delete(t.smi(i), &deleted_value));
|
|
|
|
CHECK_EQ(&t, deleted_value);
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(Delete_num_not_found) {
|
|
|
|
IdentityMapTester t;
|
|
|
|
for (int i = 0; i < 100; i++) {
|
2018-04-19 14:10:45 +00:00
|
|
|
void* deleted_value = &t;
|
|
|
|
CHECK(!t.map.Delete(t.num(i + 0.2), &deleted_value));
|
|
|
|
CHECK_EQ(&t, deleted_value);
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(GetFind_smi_0) {
|
|
|
|
IdentityMapTester t;
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.TestInsertFind(t.smi(0), t.isolate(), t.smi(1), t.heap());
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
2015-05-06 12:40:21 +00:00
|
|
|
|
|
|
|
TEST(GetFind_smi_13) {
|
|
|
|
IdentityMapTester t;
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.TestInsertFind(t.smi(13), t.isolate(), t.smi(17), t.heap());
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
TEST(GetFind_num_13) {
|
|
|
|
IdentityMapTester t;
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.TestInsertFind(t.num(13.1), t.isolate(), t.num(17.1), t.heap());
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
TEST(Delete_smi_13) {
|
|
|
|
IdentityMapTester t;
|
|
|
|
t.TestFindDelete(t.smi(13), t.isolate(), t.smi(17), t.heap());
|
|
|
|
CHECK(t.map.empty());
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(Delete_num_13) {
|
|
|
|
IdentityMapTester t;
|
|
|
|
t.TestFindDelete(t.num(13.1), t.isolate(), t.num(17.1), t.heap());
|
|
|
|
CHECK(t.map.empty());
|
|
|
|
}
|
2015-05-06 12:40:21 +00:00
|
|
|
|
|
|
|
TEST(GetFind_smi_17m) {
|
|
|
|
const int kInterval = 17;
|
|
|
|
const int kShift = 1099;
|
|
|
|
IdentityMapTester t;
|
|
|
|
|
|
|
|
for (int i = 1; i < 100; i += kInterval) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(i), reinterpret_cast<void*>(i + kShift));
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = 1; i < 100; i += kInterval) {
|
|
|
|
t.CheckFind(t.smi(i), reinterpret_cast<void*>(i + kShift));
|
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = 1; i < 100; i += kInterval) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.CheckFindOrInsert(t.smi(i), reinterpret_cast<void*>(i + kShift));
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = 1; i < 100; i++) {
|
|
|
|
void** entry = t.map.Find(t.smi(i));
|
|
|
|
if ((i % kInterval) != 1) {
|
|
|
|
CHECK_NULL(entry);
|
|
|
|
} else {
|
|
|
|
CHECK_NOT_NULL(entry);
|
|
|
|
CHECK_EQ(reinterpret_cast<void*>(i + kShift), *entry);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
TEST(Delete_smi_17m) {
|
|
|
|
const int kInterval = 17;
|
|
|
|
const int kShift = 1099;
|
|
|
|
IdentityMapTester t;
|
|
|
|
|
|
|
|
for (int i = 1; i < 100; i += kInterval) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(i), reinterpret_cast<void*>(i + kShift));
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = 1; i < 100; i += kInterval) {
|
|
|
|
t.CheckFind(t.smi(i), reinterpret_cast<void*>(i + kShift));
|
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = 1; i < 100; i += kInterval) {
|
|
|
|
t.CheckDelete(t.smi(i), reinterpret_cast<void*>(i + kShift));
|
|
|
|
for (int j = 1; j < 100; j += kInterval) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
auto entry = t.map.Find(t.smi(j));
|
2017-02-20 21:46:38 +00:00
|
|
|
if (j <= i) {
|
|
|
|
CHECK_NULL(entry);
|
|
|
|
} else {
|
|
|
|
CHECK_NOT_NULL(entry);
|
|
|
|
CHECK_EQ(reinterpret_cast<void*>(j + kShift), *entry);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2015-05-06 12:40:21 +00:00
|
|
|
|
|
|
|
TEST(GetFind_num_1000) {
|
|
|
|
const int kPrime = 137;
|
|
|
|
IdentityMapTester t;
|
|
|
|
int val1;
|
|
|
|
int val2;
|
|
|
|
|
2015-10-22 12:12:12 +00:00
|
|
|
for (int i = 0; i < 1000; i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.TestInsertFind(t.smi(i * kPrime), &val1, t.smi(i * kPrime + 1), &val2);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
TEST(Delete_num_1000) {
|
|
|
|
const int kPrime = 137;
|
|
|
|
IdentityMapTester t;
|
|
|
|
|
|
|
|
for (int i = 0; i < 1000; i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(i * kPrime), reinterpret_cast<void*>(i * kPrime));
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Delete every second value in reverse.
|
|
|
|
for (int i = 999; i >= 0; i -= 2) {
|
2018-04-19 14:10:45 +00:00
|
|
|
void* entry;
|
|
|
|
CHECK(t.map.Delete(t.smi(i * kPrime), &entry));
|
2017-02-20 21:46:38 +00:00
|
|
|
CHECK_EQ(reinterpret_cast<void*>(i * kPrime), entry);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = 0; i < 1000; i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
auto entry = t.map.Find(t.smi(i * kPrime));
|
2017-02-20 21:46:38 +00:00
|
|
|
if (i % 2) {
|
|
|
|
CHECK_NULL(entry);
|
|
|
|
} else {
|
|
|
|
CHECK_NOT_NULL(entry);
|
|
|
|
CHECK_EQ(reinterpret_cast<void*>(i * kPrime), *entry);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Delete the rest.
|
|
|
|
for (int i = 0; i < 1000; i += 2) {
|
2018-04-19 14:10:45 +00:00
|
|
|
void* entry;
|
|
|
|
CHECK(t.map.Delete(t.smi(i * kPrime), &entry));
|
2017-02-20 21:46:38 +00:00
|
|
|
CHECK_EQ(reinterpret_cast<void*>(i * kPrime), entry);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = 0; i < 1000; i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
auto entry = t.map.Find(t.smi(i * kPrime));
|
2017-02-20 21:46:38 +00:00
|
|
|
CHECK_NULL(entry);
|
|
|
|
}
|
|
|
|
}
|
2015-05-06 12:40:21 +00:00
|
|
|
|
|
|
|
TEST(GetFind_smi_gc) {
|
|
|
|
const int kKey = 33;
|
|
|
|
const int kShift = 1211;
|
|
|
|
IdentityMapTester t;
|
|
|
|
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(kKey), &t);
|
2015-05-06 12:40:21 +00:00
|
|
|
t.SimulateGCByIncrementingSmisBy(kShift);
|
|
|
|
t.CheckFind(t.smi(kKey + kShift), &t);
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.CheckFindOrInsert(t.smi(kKey + kShift), &t);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
TEST(Delete_smi_gc) {
|
|
|
|
const int kKey = 33;
|
|
|
|
const int kShift = 1211;
|
|
|
|
IdentityMapTester t;
|
|
|
|
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(kKey), &t);
|
2017-02-20 21:46:38 +00:00
|
|
|
t.SimulateGCByIncrementingSmisBy(kShift);
|
|
|
|
t.CheckDelete(t.smi(kKey + kShift), &t);
|
|
|
|
}
|
2015-05-06 12:40:21 +00:00
|
|
|
|
|
|
|
TEST(GetFind_smi_gc2) {
|
|
|
|
int kKey1 = 1;
|
|
|
|
int kKey2 = 33;
|
|
|
|
const int kShift = 1211;
|
|
|
|
IdentityMapTester t;
|
|
|
|
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(kKey1), &kKey1);
|
|
|
|
t.map.Insert(t.smi(kKey2), &kKey2);
|
2015-05-06 12:40:21 +00:00
|
|
|
t.SimulateGCByIncrementingSmisBy(kShift);
|
|
|
|
t.CheckFind(t.smi(kKey1 + kShift), &kKey1);
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.CheckFindOrInsert(t.smi(kKey1 + kShift), &kKey1);
|
2015-05-06 12:40:21 +00:00
|
|
|
t.CheckFind(t.smi(kKey2 + kShift), &kKey2);
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.CheckFindOrInsert(t.smi(kKey2 + kShift), &kKey2);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
TEST(Delete_smi_gc2) {
|
|
|
|
int kKey1 = 1;
|
|
|
|
int kKey2 = 33;
|
|
|
|
const int kShift = 1211;
|
|
|
|
IdentityMapTester t;
|
|
|
|
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(kKey1), &kKey1);
|
|
|
|
t.map.Insert(t.smi(kKey2), &kKey2);
|
2017-02-20 21:46:38 +00:00
|
|
|
t.SimulateGCByIncrementingSmisBy(kShift);
|
|
|
|
t.CheckDelete(t.smi(kKey1 + kShift), &kKey1);
|
|
|
|
t.CheckDelete(t.smi(kKey2 + kShift), &kKey2);
|
|
|
|
}
|
2015-05-06 12:40:21 +00:00
|
|
|
|
|
|
|
TEST(GetFind_smi_gc_n) {
|
|
|
|
const int kShift = 12011;
|
|
|
|
IdentityMapTester t;
|
|
|
|
int keys[12] = {1, 2, 7, 8, 15, 23,
|
|
|
|
1 + 32, 2 + 32, 7 + 32, 8 + 32, 15 + 32, 23 + 32};
|
|
|
|
// Initialize the map first.
|
|
|
|
for (size_t i = 0; i < arraysize(keys); i += 2) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.TestInsertFind(t.smi(keys[i]), &keys[i], t.smi(keys[i + 1]),
|
|
|
|
&keys[i + 1]);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
// Check the above initialization.
|
|
|
|
for (size_t i = 0; i < arraysize(keys); i++) {
|
|
|
|
t.CheckFind(t.smi(keys[i]), &keys[i]);
|
|
|
|
}
|
|
|
|
// Simulate a GC by "moving" the smis in the internal keys array.
|
|
|
|
t.SimulateGCByIncrementingSmisBy(kShift);
|
|
|
|
// Check that searching for the incremented smis finds the same values.
|
|
|
|
for (size_t i = 0; i < arraysize(keys); i++) {
|
|
|
|
t.CheckFind(t.smi(keys[i] + kShift), &keys[i]);
|
|
|
|
}
|
|
|
|
// Check that searching for the incremented smis gets the same values.
|
|
|
|
for (size_t i = 0; i < arraysize(keys); i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.CheckFindOrInsert(t.smi(keys[i] + kShift), &keys[i]);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
TEST(Delete_smi_gc_n) {
|
|
|
|
const int kShift = 12011;
|
|
|
|
IdentityMapTester t;
|
|
|
|
int keys[12] = {1, 2, 7, 8, 15, 23,
|
|
|
|
1 + 32, 2 + 32, 7 + 32, 8 + 32, 15 + 32, 23 + 32};
|
|
|
|
// Initialize the map first.
|
|
|
|
for (size_t i = 0; i < arraysize(keys); i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(keys[i]), &keys[i]);
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
// Simulate a GC by "moving" the smis in the internal keys array.
|
|
|
|
t.SimulateGCByIncrementingSmisBy(kShift);
|
|
|
|
// Check that deleting for the incremented smis finds the same values.
|
|
|
|
for (size_t i = 0; i < arraysize(keys); i++) {
|
|
|
|
t.CheckDelete(t.smi(keys[i] + kShift), &keys[i]);
|
|
|
|
}
|
|
|
|
}
|
2015-05-06 12:40:21 +00:00
|
|
|
|
|
|
|
TEST(GetFind_smi_num_gc_n) {
|
|
|
|
const int kShift = 12019;
|
|
|
|
IdentityMapTester t;
|
|
|
|
int smi_keys[] = {1, 2, 7, 15, 23};
|
|
|
|
Handle<Object> num_keys[] = {t.num(1.1), t.num(2.2), t.num(3.3), t.num(4.4),
|
|
|
|
t.num(5.5), t.num(6.6), t.num(7.7), t.num(8.8),
|
|
|
|
t.num(9.9), t.num(10.1)};
|
|
|
|
// Initialize the map first.
|
|
|
|
for (size_t i = 0; i < arraysize(smi_keys); i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(smi_keys[i]), &smi_keys[i]);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
for (size_t i = 0; i < arraysize(num_keys); i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(num_keys[i], &num_keys[i]);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
// Check the above initialization.
|
|
|
|
for (size_t i = 0; i < arraysize(smi_keys); i++) {
|
|
|
|
t.CheckFind(t.smi(smi_keys[i]), &smi_keys[i]);
|
|
|
|
}
|
|
|
|
for (size_t i = 0; i < arraysize(num_keys); i++) {
|
|
|
|
t.CheckFind(num_keys[i], &num_keys[i]);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Simulate a GC by moving SMIs.
|
|
|
|
// Ironically the SMIs "move", but the heap numbers don't!
|
|
|
|
t.SimulateGCByIncrementingSmisBy(kShift);
|
|
|
|
|
|
|
|
// Check that searching for the incremented smis finds the same values.
|
|
|
|
for (size_t i = 0; i < arraysize(smi_keys); i++) {
|
|
|
|
t.CheckFind(t.smi(smi_keys[i] + kShift), &smi_keys[i]);
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.CheckFindOrInsert(t.smi(smi_keys[i] + kShift), &smi_keys[i]);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Check that searching for the numbers finds the same values.
|
|
|
|
for (size_t i = 0; i < arraysize(num_keys); i++) {
|
|
|
|
t.CheckFind(num_keys[i], &num_keys[i]);
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.CheckFindOrInsert(num_keys[i], &num_keys[i]);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-02-20 21:46:38 +00:00
|
|
|
TEST(Delete_smi_num_gc_n) {
|
|
|
|
const int kShift = 12019;
|
|
|
|
IdentityMapTester t;
|
|
|
|
int smi_keys[] = {1, 2, 7, 15, 23};
|
|
|
|
Handle<Object> num_keys[] = {t.num(1.1), t.num(2.2), t.num(3.3), t.num(4.4),
|
|
|
|
t.num(5.5), t.num(6.6), t.num(7.7), t.num(8.8),
|
|
|
|
t.num(9.9), t.num(10.1)};
|
|
|
|
// Initialize the map first.
|
|
|
|
for (size_t i = 0; i < arraysize(smi_keys); i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(smi_keys[i]), &smi_keys[i]);
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
for (size_t i = 0; i < arraysize(num_keys); i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(num_keys[i], &num_keys[i]);
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Simulate a GC by moving SMIs.
|
|
|
|
// Ironically the SMIs "move", but the heap numbers don't!
|
|
|
|
t.SimulateGCByIncrementingSmisBy(kShift);
|
|
|
|
|
|
|
|
// Check that deleting for the incremented smis finds the same values.
|
|
|
|
for (size_t i = 0; i < arraysize(smi_keys); i++) {
|
|
|
|
t.CheckDelete(t.smi(smi_keys[i] + kShift), &smi_keys[i]);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Check that deleting the numbers finds the same values.
|
|
|
|
for (size_t i = 0; i < arraysize(num_keys); i++) {
|
|
|
|
t.CheckDelete(num_keys[i], &num_keys[i]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(Delete_smi_resizes) {
|
|
|
|
const int kKeyCount = 1024;
|
|
|
|
const int kValueOffset = 27;
|
|
|
|
IdentityMapTester t;
|
|
|
|
|
|
|
|
// Insert one element to initialize map.
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(0), reinterpret_cast<void*>(kValueOffset));
|
2017-02-20 21:46:38 +00:00
|
|
|
|
|
|
|
int initial_capacity = t.map.capacity();
|
|
|
|
CHECK_LT(initial_capacity, kKeyCount);
|
|
|
|
|
|
|
|
// Insert another kKeyCount - 1 keys.
|
|
|
|
for (int i = 1; i < kKeyCount; i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(i), reinterpret_cast<void*>(i + kValueOffset));
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Check capacity increased.
|
|
|
|
CHECK_GT(t.map.capacity(), initial_capacity);
|
|
|
|
CHECK_GE(t.map.capacity(), kKeyCount);
|
|
|
|
|
|
|
|
// Delete all the keys.
|
|
|
|
for (int i = 0; i < kKeyCount; i++) {
|
|
|
|
t.CheckDelete(t.smi(i), reinterpret_cast<void*>(i + kValueOffset));
|
|
|
|
}
|
|
|
|
|
|
|
|
// Should resize back to initial capacity.
|
|
|
|
CHECK_EQ(t.map.capacity(), initial_capacity);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(Iterator_smi_num) {
|
|
|
|
IdentityMapTester t;
|
|
|
|
int smi_keys[] = {1, 2, 7, 15, 23};
|
|
|
|
Handle<Object> num_keys[] = {t.num(1.1), t.num(2.2), t.num(3.3), t.num(4.4),
|
|
|
|
t.num(5.5), t.num(6.6), t.num(7.7), t.num(8.8),
|
|
|
|
t.num(9.9), t.num(10.1)};
|
|
|
|
// Initialize the map.
|
|
|
|
for (size_t i = 0; i < arraysize(smi_keys); i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(smi_keys[i]), reinterpret_cast<void*>(i));
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
for (size_t i = 0; i < arraysize(num_keys); i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(num_keys[i], reinterpret_cast<void*>(i + 5));
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
|
2017-03-03 17:14:40 +00:00
|
|
|
// Check iterator sees all values once.
|
2017-02-20 21:46:38 +00:00
|
|
|
std::set<intptr_t> seen;
|
|
|
|
{
|
|
|
|
IdentityMap<void*, ZoneAllocationPolicy>::IteratableScope it_scope(&t.map);
|
|
|
|
for (auto it = it_scope.begin(); it != it_scope.end(); ++it) {
|
2017-03-03 17:14:40 +00:00
|
|
|
CHECK(seen.find(reinterpret_cast<intptr_t>(**it)) == seen.end());
|
2017-02-20 21:46:38 +00:00
|
|
|
seen.insert(reinterpret_cast<intptr_t>(**it));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
for (intptr_t i = 0; i < 15; i++) {
|
|
|
|
CHECK(seen.find(i) != seen.end());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(Iterator_smi_num_gc) {
|
|
|
|
const int kShift = 16039;
|
|
|
|
IdentityMapTester t;
|
|
|
|
int smi_keys[] = {1, 2, 7, 15, 23};
|
|
|
|
Handle<Object> num_keys[] = {t.num(1.1), t.num(2.2), t.num(3.3), t.num(4.4),
|
|
|
|
t.num(5.5), t.num(6.6), t.num(7.7), t.num(8.8),
|
|
|
|
t.num(9.9), t.num(10.1)};
|
|
|
|
// Initialize the map.
|
|
|
|
for (size_t i = 0; i < arraysize(smi_keys); i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(smi_keys[i]), reinterpret_cast<void*>(i));
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
for (size_t i = 0; i < arraysize(num_keys); i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(num_keys[i], reinterpret_cast<void*>(i + 5));
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Simulate GC by moving the SMIs.
|
|
|
|
t.SimulateGCByIncrementingSmisBy(kShift);
|
|
|
|
|
|
|
|
// Check iterator sees all values.
|
|
|
|
std::set<intptr_t> seen;
|
|
|
|
{
|
|
|
|
IdentityMap<void*, ZoneAllocationPolicy>::IteratableScope it_scope(&t.map);
|
|
|
|
for (auto it = it_scope.begin(); it != it_scope.end(); ++it) {
|
2017-03-03 17:14:40 +00:00
|
|
|
CHECK(seen.find(reinterpret_cast<intptr_t>(**it)) == seen.end());
|
2017-02-20 21:46:38 +00:00
|
|
|
seen.insert(reinterpret_cast<intptr_t>(**it));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
for (intptr_t i = 0; i < 15; i++) {
|
|
|
|
CHECK(seen.find(i) != seen.end());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-03-03 17:14:40 +00:00
|
|
|
void IterateCollisionTest(int stride) {
|
|
|
|
for (int load = 15; load <= 120; load = load * 2) {
|
|
|
|
IdentityMapTester t;
|
2017-02-20 21:46:38 +00:00
|
|
|
|
2017-03-03 17:14:40 +00:00
|
|
|
{ // Add entries to the map.
|
|
|
|
HandleScope scope(t.isolate());
|
|
|
|
int next = 1;
|
|
|
|
for (int i = 0; i < load; i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(next), reinterpret_cast<void*>(next));
|
2017-03-03 17:14:40 +00:00
|
|
|
t.CheckFind(t.smi(next), reinterpret_cast<void*>(next));
|
|
|
|
next = next + stride;
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
}
|
2017-03-03 17:14:40 +00:00
|
|
|
// Iterate through the map and check we see all elements only once.
|
|
|
|
std::set<intptr_t> seen;
|
|
|
|
{
|
|
|
|
IdentityMap<void*, ZoneAllocationPolicy>::IteratableScope it_scope(
|
|
|
|
&t.map);
|
|
|
|
for (auto it = it_scope.begin(); it != it_scope.end(); ++it) {
|
|
|
|
CHECK(seen.find(reinterpret_cast<intptr_t>(**it)) == seen.end());
|
|
|
|
seen.insert(reinterpret_cast<intptr_t>(**it));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Check get and find on map.
|
|
|
|
{
|
|
|
|
HandleScope scope(t.isolate());
|
|
|
|
int next = 1;
|
|
|
|
for (int i = 0; i < load; i++) {
|
|
|
|
CHECK(seen.find(next) != seen.end());
|
|
|
|
t.CheckFind(t.smi(next), reinterpret_cast<void*>(next));
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.CheckFindOrInsert(t.smi(next), reinterpret_cast<void*>(next));
|
2017-03-03 17:14:40 +00:00
|
|
|
next = next + stride;
|
|
|
|
}
|
2017-02-20 21:46:38 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2015-05-06 12:40:21 +00:00
|
|
|
|
2017-03-03 17:14:40 +00:00
|
|
|
TEST(IterateCollisions_1) { IterateCollisionTest(1); }
|
|
|
|
TEST(IterateCollisions_2) { IterateCollisionTest(2); }
|
|
|
|
TEST(IterateCollisions_3) { IterateCollisionTest(3); }
|
|
|
|
TEST(IterateCollisions_5) { IterateCollisionTest(5); }
|
|
|
|
TEST(IterateCollisions_7) { IterateCollisionTest(7); }
|
|
|
|
|
2015-05-06 12:40:21 +00:00
|
|
|
void CollisionTest(int stride, bool rehash = false, bool resize = false) {
|
|
|
|
for (int load = 15; load <= 120; load = load * 2) {
|
|
|
|
IdentityMapTester t;
|
|
|
|
|
|
|
|
{ // Add entries to the map.
|
|
|
|
HandleScope scope(t.isolate());
|
|
|
|
int next = 1;
|
|
|
|
for (int i = 0; i < load; i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(next), reinterpret_cast<void*>(next));
|
2015-05-06 12:40:21 +00:00
|
|
|
t.CheckFind(t.smi(next), reinterpret_cast<void*>(next));
|
|
|
|
next = next + stride;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (resize) t.Resize(); // Explicit resize (internal method).
|
|
|
|
if (rehash) t.Rehash(); // Explicit rehash (internal method).
|
|
|
|
{ // Check find and get.
|
|
|
|
HandleScope scope(t.isolate());
|
|
|
|
int next = 1;
|
|
|
|
for (int i = 0; i < load; i++) {
|
|
|
|
t.CheckFind(t.smi(next), reinterpret_cast<void*>(next));
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.CheckFindOrInsert(t.smi(next), reinterpret_cast<void*>(next));
|
2015-05-06 12:40:21 +00:00
|
|
|
next = next + stride;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(Collisions_1) { CollisionTest(1); }
|
|
|
|
TEST(Collisions_2) { CollisionTest(2); }
|
|
|
|
TEST(Collisions_3) { CollisionTest(3); }
|
|
|
|
TEST(Collisions_5) { CollisionTest(5); }
|
|
|
|
TEST(Collisions_7) { CollisionTest(7); }
|
|
|
|
TEST(Resize) { CollisionTest(9, false, true); }
|
|
|
|
TEST(Rehash) { CollisionTest(11, true, false); }
|
|
|
|
|
|
|
|
TEST(ExplicitGC) {
|
|
|
|
IdentityMapTester t;
|
|
|
|
Handle<Object> num_keys[] = {t.num(2.1), t.num(2.4), t.num(3.3), t.num(4.3),
|
|
|
|
t.num(7.5), t.num(6.4), t.num(7.3), t.num(8.3),
|
|
|
|
t.num(8.9), t.num(10.4)};
|
|
|
|
|
|
|
|
// Insert some objects that should be in new space.
|
|
|
|
for (size_t i = 0; i < arraysize(num_keys); i++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(num_keys[i], &num_keys[i]);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Do an explicit, real GC.
|
2016-09-07 10:02:58 +00:00
|
|
|
t.heap()->CollectGarbage(i::NEW_SPACE, i::GarbageCollectionReason::kTesting);
|
2015-05-06 12:40:21 +00:00
|
|
|
|
|
|
|
// Check that searching for the numbers finds the same values.
|
|
|
|
for (size_t i = 0; i < arraysize(num_keys); i++) {
|
|
|
|
t.CheckFind(num_keys[i], &num_keys[i]);
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.CheckFindOrInsert(num_keys[i], &num_keys[i]);
|
2015-05-06 12:40:21 +00:00
|
|
|
}
|
|
|
|
}
|
2015-09-30 13:46:56 +00:00
|
|
|
|
2015-10-26 15:33:02 +00:00
|
|
|
TEST(CanonicalHandleScope) {
|
|
|
|
Isolate* isolate = CcTest::i_isolate();
|
|
|
|
Heap* heap = CcTest::heap();
|
|
|
|
HandleScope outer(isolate);
|
|
|
|
CanonicalHandleScope outer_canonical(isolate);
|
|
|
|
|
|
|
|
// Deduplicate smi handles.
|
2017-08-29 12:40:31 +00:00
|
|
|
std::vector<Handle<Object>> smi_handles;
|
2015-10-26 15:33:02 +00:00
|
|
|
for (int i = 0; i < 100; i++) {
|
2017-08-29 12:40:31 +00:00
|
|
|
smi_handles.push_back(Handle<Object>(Smi::FromInt(i), isolate));
|
2015-10-26 15:33:02 +00:00
|
|
|
}
|
2018-10-31 22:52:56 +00:00
|
|
|
Address* next_handle = isolate->handle_scope_data()->next;
|
2015-10-26 15:33:02 +00:00
|
|
|
for (int i = 0; i < 100; i++) {
|
|
|
|
Handle<Object> new_smi = Handle<Object>(Smi::FromInt(i), isolate);
|
|
|
|
Handle<Object> old_smi = smi_handles[i];
|
|
|
|
CHECK_EQ(new_smi.location(), old_smi.location());
|
|
|
|
}
|
|
|
|
// Check that no new handles have been allocated.
|
|
|
|
CHECK_EQ(next_handle, isolate->handle_scope_data()->next);
|
|
|
|
|
|
|
|
// Deduplicate root list items.
|
2018-07-04 09:10:05 +00:00
|
|
|
Handle<String> empty_string(ReadOnlyRoots(heap).empty_string(), isolate);
|
|
|
|
Handle<Map> free_space_map(ReadOnlyRoots(heap).free_space_map(), isolate);
|
|
|
|
Handle<Symbol> uninitialized_symbol(
|
|
|
|
ReadOnlyRoots(heap).uninitialized_symbol(), isolate);
|
2015-10-26 15:33:02 +00:00
|
|
|
CHECK_EQ(isolate->factory()->empty_string().location(),
|
|
|
|
empty_string.location());
|
|
|
|
CHECK_EQ(isolate->factory()->free_space_map().location(),
|
|
|
|
free_space_map.location());
|
|
|
|
CHECK_EQ(isolate->factory()->uninitialized_symbol().location(),
|
|
|
|
uninitialized_symbol.location());
|
|
|
|
// Check that no new handles have been allocated.
|
|
|
|
CHECK_EQ(next_handle, isolate->handle_scope_data()->next);
|
|
|
|
|
|
|
|
// Test ordinary heap objects.
|
|
|
|
Handle<HeapNumber> number1 = isolate->factory()->NewHeapNumber(3.3);
|
|
|
|
Handle<String> string1 =
|
|
|
|
isolate->factory()->NewStringFromAsciiChecked("test");
|
|
|
|
next_handle = isolate->handle_scope_data()->next;
|
2018-06-23 09:05:50 +00:00
|
|
|
Handle<HeapNumber> number2(*number1, isolate);
|
|
|
|
Handle<String> string2(*string1, isolate);
|
2015-10-26 15:33:02 +00:00
|
|
|
CHECK_EQ(number1.location(), number2.location());
|
|
|
|
CHECK_EQ(string1.location(), string2.location());
|
2017-04-26 22:16:41 +00:00
|
|
|
CcTest::CollectAllGarbage();
|
2018-06-23 09:05:50 +00:00
|
|
|
Handle<HeapNumber> number3(*number2, isolate);
|
|
|
|
Handle<String> string3(*string2, isolate);
|
2015-10-26 15:33:02 +00:00
|
|
|
CHECK_EQ(number1.location(), number3.location());
|
|
|
|
CHECK_EQ(string1.location(), string3.location());
|
|
|
|
// Check that no new handles have been allocated.
|
|
|
|
CHECK_EQ(next_handle, isolate->handle_scope_data()->next);
|
|
|
|
|
|
|
|
// Inner handle scope do not create canonical handles.
|
|
|
|
{
|
|
|
|
HandleScope inner(isolate);
|
2018-06-23 09:05:50 +00:00
|
|
|
Handle<HeapNumber> number4(*number1, isolate);
|
|
|
|
Handle<String> string4(*string1, isolate);
|
2015-10-26 15:33:02 +00:00
|
|
|
CHECK_NE(number1.location(), number4.location());
|
|
|
|
CHECK_NE(string1.location(), string4.location());
|
|
|
|
|
|
|
|
// Nested canonical scope does not conflict with outer canonical scope,
|
|
|
|
// but does not canonicalize across scopes.
|
|
|
|
CanonicalHandleScope inner_canonical(isolate);
|
2018-06-23 09:05:50 +00:00
|
|
|
Handle<HeapNumber> number5(*number4, isolate);
|
|
|
|
Handle<String> string5(*string4, isolate);
|
2015-10-26 15:33:02 +00:00
|
|
|
CHECK_NE(number4.location(), number5.location());
|
|
|
|
CHECK_NE(string4.location(), string5.location());
|
|
|
|
CHECK_NE(number1.location(), number5.location());
|
|
|
|
CHECK_NE(string1.location(), string5.location());
|
|
|
|
|
2018-06-23 09:05:50 +00:00
|
|
|
Handle<HeapNumber> number6(*number1, isolate);
|
|
|
|
Handle<String> string6(*string1, isolate);
|
2015-10-26 15:33:02 +00:00
|
|
|
CHECK_NE(number4.location(), number6.location());
|
|
|
|
CHECK_NE(string4.location(), string6.location());
|
|
|
|
CHECK_NE(number1.location(), number6.location());
|
|
|
|
CHECK_NE(string1.location(), string6.location());
|
|
|
|
CHECK_EQ(number5.location(), number6.location());
|
|
|
|
CHECK_EQ(string5.location(), string6.location());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-06-20 21:36:26 +00:00
|
|
|
TEST(GCShortCutting) {
|
2020-07-02 23:40:05 +00:00
|
|
|
if (FLAG_single_generation) return;
|
2017-10-05 11:57:49 +00:00
|
|
|
ManualGCScope manual_gc_scope;
|
2017-06-20 21:36:26 +00:00
|
|
|
IdentityMapTester t;
|
|
|
|
Isolate* isolate = CcTest::i_isolate();
|
|
|
|
Factory* factory = isolate->factory();
|
|
|
|
const int kDummyValue = 0;
|
|
|
|
|
|
|
|
for (int i = 0; i < 16; i++) {
|
|
|
|
// Insert a varying number of Smis as padding to ensure some tests straddle
|
|
|
|
// a boundary where the thin string short cutting will cause size_ to be
|
|
|
|
// greater to capacity_ if not corrected by IdentityMap
|
|
|
|
// (see crbug.com/704132).
|
|
|
|
for (int j = 0; j < i; j++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(j), reinterpret_cast<void*>(kDummyValue));
|
2017-06-20 21:36:26 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
Handle<String> thin_string =
|
|
|
|
factory->NewStringFromAsciiChecked("thin_string");
|
|
|
|
Handle<String> internalized_string =
|
|
|
|
factory->InternalizeString(thin_string);
|
2021-06-08 15:35:13 +00:00
|
|
|
DCHECK(thin_string->IsThinString());
|
2017-06-20 21:36:26 +00:00
|
|
|
DCHECK_NE(*thin_string, *internalized_string);
|
|
|
|
|
|
|
|
// Insert both keys into the map.
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(thin_string, &thin_string);
|
|
|
|
t.map.Insert(internalized_string, &internalized_string);
|
2017-06-20 21:36:26 +00:00
|
|
|
|
|
|
|
// Do an explicit, real GC, this should short-cut the thin string to point
|
|
|
|
// to the internalized string.
|
|
|
|
t.heap()->CollectGarbage(i::NEW_SPACE,
|
|
|
|
i::GarbageCollectionReason::kTesting);
|
2021-06-08 15:35:13 +00:00
|
|
|
DCHECK_IMPLIES(!FLAG_optimize_for_size,
|
2017-06-20 21:36:26 +00:00
|
|
|
*thin_string == *internalized_string);
|
|
|
|
|
|
|
|
// Check that getting the object points to one of the handles.
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
void** thin_string_entry = t.map.Find(thin_string);
|
2017-06-20 21:36:26 +00:00
|
|
|
CHECK(*thin_string_entry == &thin_string ||
|
|
|
|
*thin_string_entry == &internalized_string);
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
void** internalized_string_entry = t.map.Find(internalized_string);
|
2017-06-20 21:36:26 +00:00
|
|
|
CHECK(*internalized_string_entry == &thin_string ||
|
|
|
|
*internalized_string_entry == &internalized_string);
|
|
|
|
|
|
|
|
// Trigger resize.
|
|
|
|
for (int j = 0; j < 16; j++) {
|
Reland^4 "[serializer] Allocate during deserialization"
This relands commit 3f4e9bbe43650862884b9719494b216635f3b8c2.
which was a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
The change had an issue that embedders implementing heap tracing (e.g.
Unified Heap with Blink) could be passed an uninitialized pointer if
marking happened during deserialization of an object containing such a
pointer. Because of the 0xdeadbed0 uninitialized filler value, these
embedders would then receive the value 0xdeadbed0deadbed0 as the
'pointer', and crash on dereference.
There is, however, special handling already for null pointers in heap
tracing, also for dealing with not-yet initialized values. So, we can
make the uninitialized Smi filler be 0x00000000, and that will make such
embedded fields have a nullptr representation, making them follow the
normal uninitialized value bailouts.
In addition, it relands the following dependent changes, which are
relanding unchanged and are followup performance improvements.
Relanding them in the same change should allow for cleaner reverts
should they be needed.
This relands commit 76ad3ab59732a0ac400874b5143eb18697813cf3
[identity-map] Change resize heuristic
This relands commit 77cc96aa48a4ec2841b4c1fffd63f2a69c3f59d0
[identity-map] Cache the calculated Hash
This relands commit bee5b996aae018acf87ece5fda74e07678f3bba1
[serializer] Remove Deserializer::Initialize
This relands commit c8f73f22662a8db351123a20761d48006545af37
[serializer] Cache instance type in PostProcessNewObject
This relands commit 4e7c99abdab608ce63c7dcff5a2f891f695d529c
[identity-map] Remove double-lookups in IdentityMap
Original change's description:
> Reland^3 "[serializer] Allocate during deserialization"
>
> This is a reland of c4a062a9588b24a844a30358ac40f811d0020e0b
> which was a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
>
> Fixes TSAN errors from non-atomic writes in the deserializer. Now all
> writes are (relaxed) atomic.
>
> Original change's description:
> > Reland^2 "[serializer] Allocate during deserialization"
> >
> > This is a reland of 28a30c578c1d690b320b602d6c90e7fe1996c7f4
> > which was a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> >
> > The crashes were from calling RegisterDeserializerFinished on a null
> > Isolate pointer, for a deserializer that was never initialised
> > (specifically, ReadOnlyDeserializer when ROHeap is shared).
> >
> > Original change's description:
> > > Reland "[serializer] Allocate during deserialization"
> > >
> > > This is a reland of 5d7a29c90e85476edf44c9e9e495630cf395086a
> > >
> > > This reland shuffles around the order of checks in Heap::AllocateRawWith
> > > to not check the new space addresses until it's known that this is a new
> > > space allocation. This fixes an UBSan failure during read-only space
> > > deserialization, which happens before the new space is initialized.
> > >
> > > It also fixes some issues discovered by --stress-snapshot, around
> > > serializing ThinStrings (which are now elided as part of serialization),
> > > handle counts (I bumped the maximum handle count in that check), and
> > > clearing map transitions (the map backpointer field needed a Smi
> > > uninitialized value check).
> > >
> > > Original change's description:
> > > > [serializer] Allocate during deserialization
> > > >
> > > > This patch removes the concept of reservations and a specialized
> > > > deserializer allocator, and instead makes the deserializer allocate
> > > > directly with the Heap's Allocate method.
> > > >
> > > > The major consequence of this is that the GC can now run during
> > > > deserialization, which means that:
> > > >
> > > > a) Deserialized objects are visible to the GC, and
> > > > b) Objects that the deserializer/deserialized objects point to can
> > > > move.
> > > >
> > > > Point a) is mostly not a problem due to previous work in making
> > > > deserialized objects "GC valid", i.e. making sure that they have a valid
> > > > size before any subsequent allocation/safepoint. We now additionally
> > > > have to initialize the allocated space with a valid tagged value -- this
> > > > is a magic Smi value to keep "uninitialized" checks simple.
> > > >
> > > > Point b) is solved by Handlifying the deserializer. This involves
> > > > changing any vectors of objects into vectors of Handles, and any object
> > > > keyed map into an IdentityMap (we can't use Handles as keys because
> > > > the object's address is no longer a stable hash).
> > > >
> > > > Back-references can no longer be direct chunk offsets, so instead the
> > > > deserializer stores a Handle to each deserialized object, and the
> > > > backreference is an index into this handle array. This encoding could
> > > > be optimized in the future with e.g. a second pass over the serialized
> > > > array which emits a different bytecode for objects that are and aren't
> > > > back-referenced.
> > > >
> > > > Additionally, the slot-walk over objects to initialize them can no
> > > > longer use absolute slot offsets, as again an object may move and its
> > > > slot address would become invalid. Now, slots are walked as relative
> > > > offsets to a Handle to the object, or as absolute slots for the case of
> > > > root pointers. A concept of "slot accessor" is introduced to share the
> > > > code between these two modes, and writing the slot (including write
> > > > barriers) is abstracted into this accessor.
> > > >
> > > > Finally, the Code body walk is modified to deserialize all objects
> > > > referred to by RelocInfos before doing the RelocInfo walk itself. This
> > > > is because RelocInfoIterator uses raw pointers, so we cannot allocate
> > > > during a RelocInfo walk.
> > > >
> > > > As a drive-by, the VariableRawData bytecode is tweaked to use tagged
> > > > size rather than byte size -- the size is expected to be tagged-aligned
> > > > anyway, so now we get an extra few bits in the size encoding.
> > > >
> > > > Bug: chromium:1075999
> > > > Change-Id: I672c42f553f2669888cc5e35d692c1b8ece1845e
> > > > Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2404451
> > > > Commit-Queue: Leszek Swirski <leszeks@chromium.org>
> > > > Reviewed-by: Jakob Gruber <jgruber@chromium.org>
> > > > Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
> > > > Cr-Commit-Position: refs/heads/master@{#70229}
Bug: chromium:1075999
Change-Id: Ib514a4ef16bd02bfb60d046ecbf8fae1ead64a98
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2452689
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Ulan Degenbaev <ulan@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70366}
2020-10-07 07:31:16 +00:00
|
|
|
t.map.Insert(t.smi(j + 16), reinterpret_cast<void*>(kDummyValue));
|
2017-06-20 21:36:26 +00:00
|
|
|
}
|
|
|
|
t.map.Clear();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-09-30 13:46:56 +00:00
|
|
|
} // namespace internal
|
|
|
|
} // namespace v8
|