This allows the optimizing compiler to generate unboxed loads/stores
to fields containing double values. The double value is stored
in a reusable double object.
Unboxed loads/stores are generated for optimized code. Unoptimized code
allocates a new double on loads. To avoid performance regressions
for fields that are only written few times (e.g. only in the constructor)
I put a heuristic in place that
compares the usage count of setters and getters. Unboxed operations
are only generated if the setter is invoked a significant amount of
times (threshold is 10% of getter invocations).
The CL is so big because it changes the way LocationSummmary
is allocated: We now have a bit to generate different summaries
for optimized and unoptimized code.
R=srdjan@google.com
Review URL: https://codereview.chromium.org//99573005
git-svn-id: https://dart.googlecode.com/svn/branches/bleeding_edge/dart@31164 260f80e4-7a28-3924-810f-c04153c831b5
This bug was only exposed on ARM, but can be a problem
on other platforms as well.
When building the use intervals for a value that
is used more than once at one instruction, the use
interval was not adjusted correctly: Under certain
conditions the resulting interval can be too short.
It can occur when the first use is a writable register use,
and the second use is a normal register use.
This CL grows the use intervals to the correct size
in that case.
BUG=https://code.google.com/p/dart/issues/detail?id=11800
TEST=tests/language/regress_11800_test.dart
R=vegorov@google.com
Review URL: https://codereview.chromium.org//22412002
git-svn-id: https://dart.googlecode.com/svn/branches/bleeding_edge/dart@25810 260f80e4-7a28-3924-810f-c04153c831b5
Add profiling support to select OSR candidates and launch the compiler
for OSR, followed by entry to the function at the OSR entry point.
Implemented only on IA32 and X64. The initial implementation can be
improved in various ways --- specifically: tuning of profiling
parameters and incorporation of feedback about the actual values seen
at OSR entry.
R=fschneider@google.com
Review URL: https://codereview.chromium.org//16693006
git-svn-id: https://dart.googlecode.com/svn/branches/bleeding_edge/dart@24024 260f80e4-7a28-3924-810f-c04153c831b5
This is a first step towards fully optimizing try-catch-finally.
At a catch entry, all local variables and parameters are
expected at a fixed stack location. There is a list of
initial definitions at the catch entry block, similar to
the initial definitions at graph entry.
Inside every try-block there is a special prologue code before each
call (instruction that may throw) inside the try-block. This prologue
is similar to a parallel move instruction: It moves all locals+parameters
to the locations expected by the catch-entry block. The stack frame
is extended with the corresponding number of fixed slots right below
the normal spill slots.
Every function containing try-catch has additional compiler-
generated local variables to pass the context, the exception and
the stack trace.
Variable liveness analysis is adapted to treat locals inside try{} blocks
specially: Every call has an implicit LoadLocal of every local variable.
This CL uses a safe approximiation of liveness which can be optimized further.
Current restrictions which are planned for future CLs:
* No inlining inside try-blocks.
* No inlining of functions containing try-catch.
* No try-finally yet.
R=kmillikin@google.com
Review URL: https://codereview.chromium.org//14682020
git-svn-id: https://dart.googlecode.com/svn/branches/bleeding_edge/dart@22615 260f80e4-7a28-3924-810f-c04153c831b5
AllocationSinking pass discovers non-escaping allocations that have no input uses other than uses in the stores into its own fields.
Every environment use of such allocation is replaced by a state snapshot (MaterializeObject instruction) that describes the state of each initialized field in the object. State snapshots are computed through an additional round of load-forwarding.
Once snapshots are computed allocations are removed from the graph.
MaterializeObject instructions are not compiled into native code but produce deoptimization instructions instead that describe how object should be materialized at deoptimization.
Deoptimization instructions now follow the following format:
[mat obj #1]...[mat obj #N][ret addr][... mat arguments ...][... real frames ...]
- the prefix describes each object to materialize on deopt via kMaterializeObject instruction;
- actual values that are needed for materialization are emited as a part of bottom-most stack frame. This is done to simplify implementation: they need to be discoverable by a GC during materialization phase. At the end of deoptimization they will be removed from the stack;
- normal stack slots can refer to materialized objects via kMaterializedObjectRef instruction.
Additionally this change contains fixes in load-forwarding that are needed to guarantee that all artificial LoadField instructions inserted during AllocationSinking are correctly replaced with actual values.
Limitations of the current implementation:
- can't eliminate allocations that flow into phis but otherwise don't actually escape;
- can't sink allocations out of loops;
- allocation with type arguments are not handled.
R=regis@google.com, srdjan@google.com, zra@google.com
Review URL: https://codereview.chromium.org//14935005
git-svn-id: https://dart.googlecode.com/svn/branches/bleeding_edge/dart@22485 260f80e4-7a28-3924-810f-c04153c831b5
Remove representation from location. Presence of representation in location encoding was violating the invariant that unequal locations must be disjoint (where equality for locations is defined in terms of bitwise equality of their encoding). This could lead ParallelMoveResolver to treat XMM1 containing unboxed double as unequal location to XMM1 containing unboxed mint, which is obviously incorrect.
For similar reason eliminate kFloat32x4StackSlot and kUint32x4StackSlot distinction is eliminated and both are replaced with kQuadStackSlot. Register allocator now guarantees that no kQuadStackSlot occupies the same space as any other kDoubleStackSlot. This also shrinks optimized stack when only doubles are used (but might lead to a higher stack utilization when a mixture of doubles and quads is used).
Implement allocation of scratch Cpu and Xmm registers for ParallelMoveResolver. This also allows to remove push(eax)/pop(eax) pairs when resolving memory-memory cycles on ia32.
BUG=dart:9710
Review URL: https://codereview.chromium.org//13801014
git-svn-id: https://dart.googlecode.com/svn/branches/bleeding_edge/dart@21148 260f80e4-7a28-3924-810f-c04153c831b5
- Improve heuristic detecting interference on the back-edge. Any value live_in for the loop header can introduce interfering moves at the back edge due to control flow resolution connecting split siblings. Apply adjustments based on this heuristic at the latest point when allocating a free register;
- Don't allocate a register for constants that have no constrained uses.
R=fschneider@google.com
BUG=
Review URL: https://codereview.chromium.org//12946004
git-svn-id: https://dart.googlecode.com/svn/branches/bleeding_edge/dart@20374 260f80e4-7a28-3924-810f-c04153c831b5
This CL removes optimized access for scalarlist, and only the new TypedData classes
are optimized. I changed the runtime libraries core and math to use typedData
instead of scalarlist (Uint16List is used in StringBuffer, Uint32List by Math.random).
Instead of using LoadIndexed for internal and external arrays,
split external loads into a load of the backing store and a load
of the element.
v3 <- LoadIndexed(v1, index)
becomes
v2 <- LoadUntagged(v1, ExternalTypedData::data_offset)
v3 <- LoadIndexed(v2, index);
For this I introduce two new representations in the IL:
kUntagged (for values that hold a untagged pointer) and
kNoRepresentation (for instructions accept any input
representation)
Deoptimization does not need to know about kUntagged
since these values can never occur in the environment.
Also with this change:
* fix COMPILE_ASSERT and use it in one place.
* Cleanup IL printer output of deopt ids.
Review URL: https://codereview.chromium.org//12871010
git-svn-id: https://dart.googlecode.com/svn/branches/bleeding_edge/dart@20198 260f80e4-7a28-3924-810f-c04153c831b5
Previously we removed dead phis late, in the register allocator. This
change removes them as soon as they are discovered to be dead and packs the
phi array to squeeze out NULLs. This speeds iteration but doesn't save
space because the phi array is zone-allocated.
The PhiIterator is used everywhere to iterate phis except a few places that
need to know the phi index (e.g., SSA construction, phi elimination).
R=vegorov@google.com
Review URL: https://codereview.chromium.org//12340108
git-svn-id: https://dart.googlecode.com/svn/branches/bleeding_edge/dart@19206 260f80e4-7a28-3924-810f-c04153c831b5
This CL adds AND, OR, XOR and == operations on unboxed 64-bit integers
(aka. mints).
Unboxed mints are stored in xmm registers. Each xmm register location
has an additional bit to keep track of its value representation.
Unboxed mints are materialized on the heap on deoptmization in the same way as
unboxed doubles.
The SSE instructions used are available on all CPUs that support SSE 4.1.
Review URL: https://codereview.chromium.org//10968059
git-svn-id: https://dart.googlecode.com/svn/branches/bleeding_edge/dart@13112 260f80e4-7a28-3924-810f-c04153c831b5
fschneider@google.com