Patch to use std::move when passing AllocPolicy instances to constructors. This also fixes HashTable move constuction/assignment that previously PodAssigned the whole object including the AllocPolicy base.
Differential Revision: https://phabricator.services.mozilla.com/D36175
As discussed in the previous commit message, HashTableEntry wastes space
for common entry types. This commit reorganizes the internal storage to
store all the hashes packed together, followed by all the entries, which
eliminates the aforementioned waste.
HashTableEntry's data layout currently wastes a fair amount of space due
to ABI-mandated padding. For instance, HashTableEntry<T*> on a 64-bit
platform looks like:
class HashTableEntry {
HashNumber mKeyHash;
// Four bytes of wasted space here to pad mValueData to the correct place.
unsigned char mValueData[sizeof(T*)];
};
This wasted space means that sets of pointers backed by
mozilla::HashTable waste a quarter of their entry storage space. Maps
of pointers to pointers waste a sixth of their entry storage space.
We'd like to fix this by packing all the cached hashes together,
followed by all the hash table entries.
As a first step to laying out the hash table storage differently, we
have to make HashTable not access entries directly, but go through an
abstraction that represents the key and the entry. We call this
abstraction "slots". This commit is similar to the change done for
PLDHashTable previously.
Parts of HashTable still work in terms of Entry; the creation and
destruction of tables was not worth changing here. We'll address that
in the next commit.
We do this to make our lives easier in later patches; this check
guarantees that we don't need padding between the block of cached hash
values and the block of entries immediately after it.
As discussed in the previous commit message, HashTableEntry wastes space
for common entry types. This commit reorganizes the internal storage to
store all the hashes packed together, followed by all the entries, which
eliminates the aforementioned waste.
HashTableEntry's data layout currently wastes a fair amount of space due
to ABI-mandated padding. For instance, HashTableEntry<T*> on a 64-bit
platform looks like:
class HashTableEntry {
HashNumber mKeyHash;
// Four bytes of wasted space here to pad mValueData to the correct place.
unsigned char mValueData[sizeof(T*)];
};
This wasted space means that sets of pointers backed by
mozilla::HashTable waste a quarter of their entry storage space. Maps
of pointers to pointers waste a sixth of their entry storage space.
We'd like to fix this by packing all the cached hashes together,
followed by all the hash table entries.
As a first step to laying out the hash table storage differently, we
have to make HashTable not access entries directly, but go through an
abstraction that represents the key and the entry. We call this
abstraction "slots". This commit is similar to the change done for
PLDHashTable previously.
Parts of HashTable still work in terms of Entry; the creation and
destruction of tables was not worth changing here. We'll address that
in the next commit.
We do this to make our lives easier in later patches; this check
guarantees that we don't need padding between the block of cached hash
values and the block of entries immediately after it.
Currently lookupOrAdd() will allocate if the table has no storage. But it
doesn't report an error if the allocation fails, which can cause problems.
This patch changes things so that lookupOrAdd() doesn't allocate when the table
has no storage. Instead, it returns an AddPtr that is not *valid* (its mTable
is empty) but it is *live*, and can be used in add(), whereupon the allocation
will occur.
The patch also makes Ptr::isValid() and AddPtr::isValid() non-public, because
the valid vs. live distinction is non-obvious and best kept hidden within the
classes.
--HG--
extra : rebase_source : 95d58725d92cc83332e27a61f98fa61185440e26
Because they each only have a single call site, and the "too many items
removed?" test is already encapsulated within rehashIfOverloaded().
--HG--
extra : rebase_source : f6550c483477c2839764e7f657b542c24c82b1e8
infallibleRehashIfOverloaded() actually matches what it does.
Also, the patch removes the `overloaded()` test within the function, because
`rehashIfOverloaded()` has the same test.
--HG--
extra : rebase_source : 2799e623fbdd4b8bb82f8b83166752e02a0f9202
Specifically:
- checkOverloaded -> rehashIfOverloaded
- checkUnderloaded -> shrinkIfUnderloaded
- checkOverRemoved -> rehashIfOverRemoved
Because I've always found that the `check` prefix doesn't clearly explain that
the table might be changed,
And:
- shouldCompressTable -> overRemoved
Because that matches `overloaded` and `underloaded`.
--HG--
extra : rebase_source : 56e9edd012f4a400ac366895d05ea93fb09ec6b3
Because that's a more accurate description of what it does -- it finds free
*and* removed entries.
--HG--
extra : rebase_source : 72b049f44c61a7406d9691a9d9b6405dd696848c
Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew().
This removes the need for init() and initialized(), and matches how
PLDHashTable/nsTHashtable work. It also removes the need for init() functions
in a lot of types that are built on top of mozilla::Hash{Map,Set}.
Pros:
- No need for init() calls and subsequent checks.
- No memory allocated for empty tables, which are not that uncommon.
Cons:
- An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(),
because the existing checkOverloaded() can handle it.
Specifics:
- Construction now can take a length parameter.
- init() is removed. Explicit length-setting, when necessary, now occurs in the
constructors.
- initialized() is removed.
- capacity() now returns zero when the entry storage is absent.
- lookupForAdd() is no longer `const`, because it can instantiate the storage,
which requires modifications.
- lookupForAdd() can now return an invalid AddPtr in two cases:
- old: hashing failure (due to OOM in the hasher)
- new: OOM while instantiating entry storage
The existing failure handling paths for the old case work for the new case.
- clear(), finish(), and clearAndShrink() are replaced by clear(), compact(),
and reserve(). The old compactIfUnderloaded() is also removed.
- Capacity computation code is now in its own functions, bestCapacity() and
hashShift(). setTableSizeLog2() is removed.
- uint32_t is used throughout for capacities, instead of size_t, for
consistency with other similar values.
- changeTableSize() now takes a capacity instead of a deltaLog2, and it can now
handle !mTable.
Measurements:
- Total source code size is reduced by over 900 lines. Also, lots of existing
lines got shorter (i.e. two checks were reduced to one).
- Executable size barely changed, down by 2 KiB on Linux64. The extra branches
are compensated for by the lack of init() calls.
- Speed changed negligibly. The instruction count for Bench_Cpp_MozHash
increased from 2.84 billion to 2.89 billion but any execution time change was
well below noise.
Hash{Map,Set}::putNew() can be used on a table that has had elements removed,
despite some comments to the contrary.
This patch fixes those comments. It also clarifies when putNewInfallible() can
be used.
This patch also removes the !isRemoved() assertion in findFreeEntry(), which is
confusing -- !isLive() would be more precise, but also obvious from the
surrounding code.
MozReview-Commit-ID: q4qwKGBsHx
--HG--
extra : rebase_source : 94331879f9a1484159e030de93bd0ab222b54385
Because it's quite strange, badly named, not that useful, and barely used.
Also remove WeakMap::lookupWithDefault(), which is similar, but not used at
all.
MozReview-Commit-ID: IhIl4hQ73U1
--HG--
extra : rebase_source : 7da237a56391836ca5d056248f18bd5e2d8b1564
Because (a) it's kinda weird, and (b) only used in a single test, where it can
be easily replaced with a vanilla add().
MozReview-Commit-ID: L4RoxFb7yGG
--HG--
extra : rebase_source : 515a5ede5d417686907345ad9069c6a41669dd17
It's identical to mozilla::CStringHasher.
Also fix a comment at the top of HashTable.h about CStringHasher.
--HG--
extra : rebase_source : 92176c4f6ea8041f309764b4ce0271a494853a7b
I bet nobody has used them in years. The use of ad hoc profiling (printfs plus
post-processing with a tool like https://github.com/nnethercote/counts) is
generally a better approach.
Bug 1179657 did likewise for PLDHashTable.
--HG--
extra : rebase_source : 57ab26c62f2f1eff2eec827564a3c3ff0af12f01
And add comments delimiting different operation groups. Together these changes
make Hash{Set,Map} much easier to navigate.
--HG--
extra : rebase_source : 21a7da92f898b318a24085b5f0885da85aa3afd1
Again inspired by PLDHashTable, this makes things a bit clearer than passing 0
or sCollisionBit.
It also guarantees more strongly that we'll end up with appropriately
specialized code for the Add vs. NonAdd cases. (That guarantee isn't currently
needed because the compiler inlines things sufficiently anyway, but it can't
hurt.)
--HG--
extra : rebase_source : c2341d7df85f44a00bd1b52eac0e08ac63525196
There are two improvements here.
- When we're just doing a lookup (i.e. aCollisionBit==0), we don't need to
do any special handling of removed entries. (Inlining means that the removed
entry code is entirely removed for lookups.)
- When we're doing an insertion (i.e. aCollisionBit==sCollisionBit), we now
stop adding collision markings once we find a removed entry, because they're
unnecessary after that point.
This change brings the code in alignment with PLDHashTable::SearchTable().
--HG--
extra : rebase_source : 8d6e3a88ac24d0e71a2576997face3bea971c71f
This patch does the following:
- Adds a bunch of useful high-level info at the top of the file, making the
types easier to use for newcomers.
- Adds a comment to every Hash{Set,Map} method that lacked one.
- Tweaks lots of existing comments for clarity.
- Removes comments in Hash{Set,Map} referring to details within HashTable. That
dependence is now covered in the top-level comment.
- Tweaks the signatures in hash policy classes to be more consistent.
--HG--
extra : rebase_source : ab9d6ae38b4b73e12151a228107ad3f3f12e36b3
These basically duplicate the existing Range and Enum classes, but use more
familiar terminology, similar to the iterators we have for
PLDHashTable/nsTHashtable:
- Hash{Set,Map}::all() Hash{Set,Map}::iter()
- Enum constructor Hash{Set,Map}::modIter()
- Range::front() Iterator::get()
- Range::popFront() Iterator::next()
- Range::empty() Iterator::done()
- Enum::mutableFront() ModIterator::getMutable()
- Enum::removeFront() ModIterator::remove()
- Enum::rekeyFront() ModIterator::rekey()
The next patch will reduce the amount of code duplication.
--HG--
extra : rebase_source : 5787756b144bbaea98f381d5a128cb42edb82c41
It's only used by InlineTable::Range, and can be avoided by using
mozilla::Maybe.
This also means Range::mTable can be changed from a pointer to a reference,
like Enum::mTable.
--HG--
extra : rebase_source : 35eb5b0bede361c2c765652e1e4c952f128702ef
In PLDHashTable the equivalent functions have a "Shallow" prefix, which makes
it clear that they don't measure things hanging off the table. This patch makes
mozilla::Hash{Set,Map} do likewise.
MozReview-Commit-ID: 3kwCJynhW7d
--HG--
extra : rebase_source : 9c03d11f376a9fd4cfd5cfcdc0c446c00633b210
This patch converts HashTable.h from SpiderMonkey style to Gecko style.
It was created by first running `./mach clang-format -p mfbt/HashTable.h` to
fix the whitespace, and then doing manual patch-ups, the most significant of
which were:
- Adding braces around single-statement if/else/for/while blocks.
- Renaming class members to `mFoo` form and arguments to `aFoo` form.
- Converting `typedef Old New` to `using New = Old` throughout.
MozReview-Commit-ID: LJ1emPyuAjK
The main change is that the patch copies js/public/HashTable.h to
mfbt/HashTable.h, and then changes it as follows.
- Changes `js` namespaces to `mozilla` (and removes some now-unnecessary
`mozilla::` qualifiers).
- Changes the default AllocPolicy from the SpiderMonkey-specific
`TempAllocPolicy` to the generic `MallocAllocPolicy`.
- Adds `#include "AllocPolicy.h"` (like mfbt/Vector.h).
- Changes `JS_DEBUG` use to `DEBUG`.
- Minor comment updates, as necessary.
js/public/HashTable.h is now tiny, holding just a few renamings of things from
the `mozilla` namespace into the `js` namespace to minimize churn elsewhere.
(Those renamings keep `TempAllocPolicy` as the default AllocPolicy for
js::Hash{Set,Map}.)
Also, various template specializations had to be moved from the `js` namespace
to the `mozilla` namespace to avoid compile errors.
MozReview-Commit-ID: GS9Qn9YeYDA
--HG--
rename : js/public/HashTable.h => mfbt/HashTable.h
Sylvestre Ledru
Jon Coppeard
arthur.iakab
Coroiu Cristina
Steve Fink
Nicholas Nethercote
Nathan Froyd
Cosmin Sabou