| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
|
| |
Also added some "white-box" unit tests and sprinkled in a bit of
documentation. The case that went wrong is covered by the test
rangeSplitting_1: before the fix, the new interval would have
two ranges: [66-64],[70-71]. The first range is invalid and should not
be there at all.
Change-Id: If0742f4e6a96d98ea5d696f95126886ba66f92bb
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This reduces the work for the dominator tree/frontier calculations,
because there are less blocks to consider. All blocks inserted by
splitting the critical edges, have (by definition) no effect on the
dominator calculations. However, the immediate dominators for all new
blocks needs to be added, because this information is used by the block
scheduling.
This change reduces memory/time usage during optimization passes,
especially when processing excessively big switch statements.
Change-Id: Ia69882e9dabdddffa1c98b1079012d8d988e1e8f
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Changes to datastructures and more re-using of locally used temporary
vectors. For the test regress-74474-002.js this lowers the total
allocated memory from 1.98GB to 158MB. Thse peak memory usage stays at
75MB. There is no functional change.
This should give a modest performance improvement which mainly depends
on the speed of malloc()/free().
Change-Id: I1877c1903e59a33ee79ff2b801ef6f2c1cee30a6
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
| |
The basic block scheduling uses this information to place loops. When
the immediate dominator information is invalid, the scheduling can be
sub-optimal, or will sometimes forget to schedule some blocks.
Change-Id: Iaeb45f2b757b676310be25a658ceadc07d5722ec
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When doing edge resolving, too many spills were generated, and the
dependency tracking of moves was not complete. Now we only insert spills
that are caused by phi-nodes (because any other spill would be generated
at the point a variable was defined). However, there can still be
multiple dependencies between the moves generated by the edge resolving.
Instead of only checking the first dependency, all of them are tracked.
The bug report was a case where an unneccesary spill was generated, that
got tracked, but "suppressed" the other (valid!) dependent move. The
randomness was caused by the hash seeding of QHash.
Task-number: QTBUG-35840
Change-Id: Ifbc3c8fc13de53c46a8b5859721b2497189921a3
Reviewed-by: Fawzi Mohamed <fawzi.mohamed@digia.com>
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Changes the dominator frontier storage from a set of basic-blocks for
every basic-block to a BasicBlockSet for every basic-block. This new
class stores a maximum of 8 nodes in a vector, and switches to a bit
vector when going beyond 8 nodes.
This is important in two cases: most basic-blocks have 2-3 nodes in the
frontier, and an array is faster than a set in these cases. The few
cases where the frontier goes beyond 8 nodes, is when a switch
statement is used with lots of cases that all fall-through.
On regress-74474-003.js this reduces peak memory usage from 1.68G
to 60M. The switch statement in this test results in 27000 basic-blocks.
Change-Id: I42646522ba9f8642d42a5d70fc6b760bb47ae69f
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
| |
Calculation of all the children of nodes in the dominator tree is now
calculated as a local variable right before computing the dominator
frontier. The effect is that they are not retained after their only use.
Change-Id: I83c962c691b78cb767708eb04cf30d3b7a760deb
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
|
| |
Both type inference and the optimization pass do not do anything with
unconditional jumps. So, instead of adding them to the worklist and
later on removing them again, it’s faster to never add them in the first
place.
Change-Id: Ib81d43e9ea6df2b1a70e9dd1e9b9c29cb6d345d2
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
| |
Replace hashes from basic-block to basic-block with vectors that
associate basic-block index to basic-block index.
Change-Id: I834ea3d825e4d2b02c075b1b0f080f5a65f41317
Reviewed-by: Fawzi Mohamed <fawzi.mohamed@digia.com>
|
|
|
|
|
|
|
| |
Also fixed some comments.
Change-Id: I4aedff84bdbf8de248025bc48805a859704bdd8a
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
| |
This makes time- and memory-complexity a lot better when compiling
big JavaScript functions.
Change-Id: I2a7cb9b5979844254747fa5cf7355cca0b113904
Reviewed-by: Fawzi Mohamed <fawzi.mohamed@digia.com>
|
|
|
|
|
|
|
|
|
| |
Replace the recursive calls and subsequent clean-ups by pushing actions
on a to-do stack, and processing that stack in a loop.
Change-Id: I83536e88d400592b6e9f5fda3d795e41711a131a
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Reviewed-by: Fawzi Mohamed <fawzi.mohamed@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
As a follow-up to the previous commit, this patch cleans up the data structures
used to track dependencies of QML binding expressions and functions to context
and scope properties, determined at compile time.
Instead of "collecting" these depending properties upfront (codegen time), we
propagate the information that a property is a context or scope property into
the IR at codegen time and later in the isel collect these properties and their
notify signal index in a hash in the IR functions. The CompileData structure
generator then can read these hashes directly when writing out the dependency
information.
Change-Id: I32134706e2d24bf63d1b1abad0259ab072460173
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
| |
Change-Id: I5cb0c7798d0e530f3137710bf0e723bd7b64dc89
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
E.g.:
a | 0
b & 0xffffffff
These operations force the operands to be converted to int32 without
changing their value. At this point we already added convert calls to
the IR, so we can safely get rid of these operations.
Change-Id: Ic4d3b989e13439eccd2c878fa7bf5030acae7630
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
| |
Fixes a crash in octane.
Change-Id: Ib72ac0b7a2941230a87543f30fcf7e55d7094886
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
|
|
| |
We can't do a fast property index based access on them, due to the inability to
read individual fields from the original object (i.e. the logic in
QQmlValueTypeWrapper). However what we can determine and propagate is the type
information of the individual properties, i.e. that the x and y properties of a
QPointF are always doubles.
Change-Id: Iee71ece2117294b7bc0b93deb0a77d7c51148b11
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
|
|
| |
* Resolve lookups in namespaces at compile time and instruct the SSA optimizer
to eliminate reads from the namespace (QQmlTypeWrapper) if possible. For example
access to attached properties of types (i.e. MyNameSpace.ListView.isCurrentItem)
requires neither reading the namespace nor the type.
* Add support for accelerated lookup of attached properties
Change-Id: Ib0b66404ed7e70e1d4a46a1ac8218743a4cc8608
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
| |
With this patch we determine the meta-object of singletons, propagate it into
the IR and load them separately using a dedicated run-time function. In
addition enums in singletons and QML types are resolved at compile time.
Change-Id: I01ce1288391b476d1c9af669cb2987a44c885703
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
| |
It is technically redundant to the Member::property field.
Change-Id: If0ee35b2c94a2c9373784d36a1f8dfe8ad7dcfb3
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Propagate QObject properties in member expressions across temporaries
as part of the type interference SSA pass. This replaces the earlier
attempt to resolving QObject properties in fieldMemberExpression()
in the codegen, but it was incomplete and now things like the following
are fully resolved:
var tmp = blah.somePropertyThatReturnsAQQuickItem; <-- QQuickItem property return type propagated into tmp
var width = tmp.width; <-- and picked up here again to resolve the index of width instead of by name
With this patch Temp gets a helper structure with a function pointer,
initialized to aid the resolution of properties in Qt meta objects. This
structure is propagated into the temps until it reaches the next member
expression that uses the temp. Similarly QObjectType is added as IR type, next
to VarType.
The resolution inside the SSA type interference pass also requires passing
through the QQmlEngine from the upper caller levels, in order to resolve the
property type to a potential QMetaObject property.
Change-Id: I14c98fa455db57603da46613ce49c174d0944291
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
|
| |
Assert failed for cases where the node’s ancestor with lowest
semi-dominator number was not the same as the parent. The test case
exemplifies this.
Task-number: QTBUG-34792
Change-Id: Ie6847b22a27211801bff7479bfcbfaf329c6005a
Reviewed-by: Fawzi Mohamed <fawzi.mohamed@digia.com>
|
|
|
|
|
|
|
|
|
|
|
| |
Changed three recursive routines to worklist-based iterative ones. This
not only speeds up the dominator frontier calculation, but also prevents
the algorithm to run out of stack space.
This is a partial fix for QTBUG-34047.
Change-Id: Ife8dc35724d50408ad356e1621884bdb82db9626
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The previous version left unreachable cycles untouched. For example in:
function f() {
if (false)
while (true) { doSomething(); }
anotherThing();
}
The edge to the then-part would be removed, but the loop itself would
not be removed. This resulted in the basic-block scheduler choking when
hitting the block with the anotherThing() call, because it wants to have
all blocks from incoming edges resolved first.
Task-number: QTBUG-34776
Change-Id: I5b3a79140e6058c4ade4ec7687c1a795f1a74f97
Reviewed-by: Fawzi Mohamed <fawzi.mohamed@digia.com>
Reviewed-by: Mitch Curtis <mitch.curtis@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
| |
The engine used to round, but that is inconsistent with ECMAScript's way of
converting doubles to integers by truncation.
With this patch we can also enable the propagation of integer type information
into the IR, but we have to be careful not to utilize it when writing
properties.
Change-Id: I04af4879ba5131349eca2eeff2b27f4598f5267b
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
| |
Change-Id: Ic0492fbe31a1e134674bc6c20381f735dd6d5b7a
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
... by mapping the property type to the IR type if possible. In an expression
like parent.width * 0.5 this avoids a to-double conversion for parent.width but
instead we can rely on the result of the property read for the width property
to always be a double.
Unfortunately integer propertyes are currently not propagated, because upon
assignment from a double we would do the ECMAScript compliant truncation while
QML actually expects a round to happen. This needs to be solved separately.
Change-Id: I9c8f58416201d406e6e11d157cae12a686b774e5
Reviewed-by: Erik Verbruggen <erik.verbruggen@digia.com>
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Start the work to remove c++ exceptions from our JS
exception handling. Rather rely on engine->hasException.
Check the flag after we return from any runtime call in the
JIT.
Implement new try/catch handling code in qv4codegen and
for the JIT that doesn't rely on exceptions. As an added
bonus, we can remove the Try statement in the IR.
Change-Id: Ic95addd6ae03371c43c47e04cac26afdce23a061
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
| |
Change-Id: I47507a4d7d1b429b9c43ed3a7822079efe577327
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
| |
The unsigned right shift operations can use a signed int as input, unifying
the handling with the other shift operations. The only difference now is
in the type of the return value.
Change-Id: Ia9b83568951d1c1c77322d07cd734e2c37d34573
Reviewed-by: Erik Verbruggen <erik.verbruggen@digia.com>
|
|
|
|
|
|
|
|
|
| |
We don't need the right side of the shift operation as uint.
Converting it to int is cheaper and more then enough, as all
but the lowest 5 bits are ignored anyway.
Change-Id: I8833e6cc4e565b8bd1e35a22250e03a9b34938df
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Do not generate jump instructions when the target immediately follows
the current basic block, even if there are intermediate jumps in between
as long as they jump to the same basic block. In the IR snippet below,
no jumps will be generated at all.
…
L8: goto L6;
L12: goto L6;
L6: goto L4;
L11: goto L4;
L4: goto L2;
L10: goto L2;
L2: ….
Before this change, the gotos in L8, L6, and L2 were still generated.
Change-Id: I718ed0d41c603a6905f2279b782cd9e9cafb7d55
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
| |
This makes the code go quite a bit faster (saves ~7-8% of the
total amount of instructions executed when running crypto.js
Change-Id: I6b3bd08eca98b45593262e2fc6e0ce5056257e76
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
| |
Change-Id: If00a108fb107d331478dd36ad7feae4c4521c2ae
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
| |
Change-Id: I889e760f75b485a28e1f2a2c26b2337ae9bfafac
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When a phi-node couldn't be fully typed (e.g., when one of the temps
was not yet typed), VarType was assumed. When a circular dependency
between two phi-nodes occurred, like with a condition inside a loop,
then depending on the ordering of the work-list, the two phi-nodes
could start oscillating between VarType and the correct type.
The fix is to check if one of the temps is not fully typed, and if so,
assume whatever we currently have as the result and have the statement
re-scheduled. Full typing will occur when the temp with the missing
type information is typed.
Change-Id: I950d81fe7fa8272cb37f7eea5b88092d1eb4817e
Reviewed-by: Tor Arne Vestbø <tor.arne.vestbo@digia.com>
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This gives slightly better results for inplace increment/decrement,
where the operand has to be converted to a number. When that operand is
known to be a number, this conversion can be removed. By doing the
optimizations after typing, these conversions will not only turn into
assignments, but they will be propagated through the IR. That saves
a copy/move, which, in turn, will lighten the work for the register
allocator and the instruction selection.
Change-Id: If76575a71fbcb2d810b94060e61b4364bdaaa065
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
| |
Inplace operations are expanded when building the IR, so the neither the
IR, nor the instruction selection backends or runtime need to handle
them.
Change-Id: Id01f9544e137dd52364cf2ed2c10931c31ddfff3
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
| |
Use QV4_SHOW_ASM for the generated assembly from the JIT, and use
QV4_SHOW_IR to get dumps of the IR.
Change-Id: Id85d3d6c87b47088c312475a7c737d54c58c7791
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
| |
Also clean up a few other direct uses of Value
Change-Id: Ie27d42c1b31b9e6d16d0a60071cb5e4e1c5b9e8b
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
| |
Because "targetTemp" and "sourceTemp" is clearer than, say, t1 and t2.
Change-Id: I5195c75f1958a3bea3c370924685ff6ba9c9c515
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
|
|
|
| |
When resolving conflicting register use between basic blocks, only
insert the resolving moves into the successor when it has one incoming
edge. Because of the absence of critical edges, this implies that it is
also save to insert those moves into the predecessor block if there is
more than one incoming edge (the predecessor will only have one outgoing
edge).
Change-Id: I83c41b4ca86946d3aa09619f20ddab3e692136f2
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
| |
But keep the register allocator disabled for now.
Change-Id: I475835ec35ef31d76e084788a754c00b1f8fdda6
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
| |
The ObjectType was a misnomer: it was used to indicate that the
expression could have multiple types, or that the type could not be
inferred statically.
Change-Id: Ic48a0cd1dd7ae7bfafd361e0c9792ab161417039
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
|
|
|
|
| |
Add scheduling for moves generated by removing phi nodes by re-using the
MoveMapping class from the register allocator. This change applies to
both the JIT when no register allocator is used, and the interpreter.
Change-Id: I38eac5b13759c7790132d1ef07fa17fcb53187e3
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
| |
QtAlgorithms is getting deprecated,
see http://www.mail-archive.com/development@qt-project.org/msg01603.html
Change-Id: I8fa7d0186cc8f0ba562695974829e37f1eb87f2f
Reviewed-by: Gunnar Sletta <gunnar.sletta@digia.com>
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
|
|
|
|
|
| |
Change-Id: Ie1de760824d8927cf10cbc1e02145f40d812e8ee
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
| |
Change-Id: Ibf14d8835517981eda41ee768446da49965248be
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
|
|
| |
This converts all methods in qv4runtime_p.h to not
use raw values in arguments anymore.
The conversion of return values will be done in a separate
commit.
Change-Id: Ie6e8f3bed459d09cb831f7f87920b7eada161502
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|
|
|
|
|
|
|
|
|
| |
Statically calculate binary expressions which have two constants as
operands. Currently only for add/subtract/multiply/divide and
comparisons.
Change-Id: Ia8c7222d45bbba956025fe349fc1494015a3e74f
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
|