- Make constants more interesting.
- Add an addition to be done after the inlined call in the try-block.
- On command line, have a bit more output.
- New alternative that deopts from unoptimized code.
BUG=
R=jarin
Review-Url: https://codereview.chromium.org/2285743002
Cr-Commit-Position: refs/heads/master@{#38974}
These tests were spliced out of changelist 2216353002 and extended.
BUG=
Review-Url: https://codereview.chromium.org/2245263003
Cr-Commit-Position: refs/heads/master@{#38906}
Reason for revert:
Octane/Mandreel aborts with an exception now:
TypeError: __FUNCTION_TABLE__[(r2 >> 2)] is not a function
Original issue's description:
> [turbofan] Insert dummy values when changing from None type.
>
> Currently we choose the MachineRepresentation::kNone representation for
> values of Type::None, and when converting values from the kNone representation
> we use "impossible" conversions that will crash at runtime. This
> assumes that the impossible conversions should never be hit (the only
> way to produce the impossible values is to perform an always-failing
> runtime check on a value, such as Smi-checking a string). Note that
> this assumes that the runtime check is executed before the impossible
> convesrion.
>
> Introducing BitwiseOr type feedback broke this in two ways:
>
> - we always pick Word32 representation for bitwise-or, so the
> impossible conversion does not trigger (it only triggers with
> None representation), and we could end up with unsupported
> conversions from Word32.
>
> - even if we inserted impossible conversions, they are pure conversions.
> Since untagging, bitwise-or operations are also pure, we could hoist
> all these before the smi check of the inputs and we could hit the
> impossible conversions before we get to the smi check.
>
> This CL addresses this by just providing dummy values for conversions
> from the Type::None type. It also removes the impossible-to-* conversions.
>
> BUG=chromium:638132
>
> Committed: https://crrev.com/c83b21ab755f1420b6da85b3ff43d7e96ead9bbe
> Cr-Commit-Position: refs/heads/master@{#38883}
TBR=mstarzinger@chromium.org,jarin@chromium.org
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=chromium:638132
Review-Url: https://codereview.chromium.org/2280613002
Cr-Commit-Position: refs/heads/master@{#38893}
Currently we choose the MachineRepresentation::kNone representation for
values of Type::None, and when converting values from the kNone representation
we use "impossible" conversions that will crash at runtime. This
assumes that the impossible conversions should never be hit (the only
way to produce the impossible values is to perform an always-failing
runtime check on a value, such as Smi-checking a string). Note that
this assumes that the runtime check is executed before the impossible
convesrion.
Introducing BitwiseOr type feedback broke this in two ways:
- we always pick Word32 representation for bitwise-or, so the
impossible conversion does not trigger (it only triggers with
None representation), and we could end up with unsupported
conversions from Word32.
- even if we inserted impossible conversions, they are pure conversions.
Since untagging, bitwise-or operations are also pure, we could hoist
all these before the smi check of the inputs and we could hit the
impossible conversions before we get to the smi check.
This CL addresses this by just providing dummy values for conversions
from the Type::None type. It also removes the impossible-to-* conversions.
BUG=chromium:638132
Review-Url: https://codereview.chromium.org/2266823002
Cr-Commit-Position: refs/heads/master@{#38883}
Unfortunately, I was unable to produce a repro without asm.js. In normal
JavaScript, the bounds check renaming saves us.
I have not done anything about the index variable aliasing and handling
of differently sized elements yet!
BUG=chromium:639210, v8:5266
Review-Url: https://codereview.chromium.org/2270793004
Cr-Commit-Position: refs/heads/master@{#38874}
Unify the representation selection rules for NumberAdd/Subtract and
SpeculativeNumberAdd/Subtract wrt. Int32Add/Sub selection. We can
safely use Int32Add/Sub as long as the inputs are in the safe additive
integer range and the output is either truncated to Word32 or provably
in Signed32 or Unsigned32 range.
R=jarin@chromium.org
Review-Url: https://codereview.chromium.org/2253293005
Cr-Commit-Position: refs/heads/master@{#38746}
At the moment, two NumberConstant nodes get different type even if their
value is the same because we always allocate a new heap number for
each number constant. This can lead to replacing a node with a node of
disjoint type in value numbering, which can result in incorrect code
down the line because of inconsistent types.
This fix makes sure that we only replace a node with a sub-type
node. Once we introduce a proper type for number constants, we can
move back to the intersection typing in value numbering.
Unfortunately, it is quite hard to write a repro for this because we cache NumberConstant nodes. We only throw away cached values that have too many conflicts (>5), so the test has to contain values that fall into the same bucket. That's where the magic floating point numbers in the test come from (they have the same low 8-bits of their hashes).
BUG=chromium:633497
Review-Url: https://codereview.chromium.org/2251833002
Cr-Commit-Position: refs/heads/master@{#38675}
This adds a very first version of inlined Array.prototype.pop into
TurboFan optimized code. We currently limit the inlining to fast
object or smi elements, until the unclear situation around hole NaNs
is resolved and we have a clear semantics inside the compiler.
It's also probably overly defensive in when it's safe to inline
the call to Array.prototype.pop, but we can always extend that
later once we have sufficient trust in the implementation and see
an actual need to extend it.
BUG=v8:2229,v8:3952,v8:5267
R=epertoso@chromium.org
Review-Url: https://codereview.chromium.org/2239703002
Cr-Commit-Position: refs/heads/master@{#38578}
This switches the interface of the runtime profiler to use frames as
opposed to functions for performing on-stack replacement. Requests for
such replacements need to target a specific frame. This will enable us
to activate bytecode as well as baseline code for the same function.
The existing %OptimizeOsr runtime function also had to adapted and now
takes an optional stack depth to target a specific stack frame.
R=bmeurer@chromium.org
BUG=v8:4764
Review-Url: https://codereview.chromium.org/2230783004
Cr-Commit-Position: refs/heads/master@{#38548}
We now deopt when the lhs of a mod is negative and the rhs is 1 too (previously, we erroneusly returned 0 instead of -0).
BUG=v8:5278
R=bmeurer@chromium.org
Review-Url: https://codereview.chromium.org/2233713002
Cr-Commit-Position: refs/heads/master@{#38525}
If we infer loop variable bounds, we need to insert a type rename node
(sigma) to make sure that simplified lowering can choose representations
consistently.
Review-Url: https://codereview.chromium.org/2222513002
Cr-Commit-Position: refs/heads/master@{#38391}
This introduces a bunch of new tests that test various aspects of
accessor inlining in TurboFan (without the actual inlining), and does
the appropriate fixes to the AstGraphBuilder. The actual inlining CL
will land separately (so we don't need to revert the tests and fixes
if the accessor CL has to be reverted).
R=jarin@chromium.org
Review-Url: https://codereview.chromium.org/2197913002
Cr-Commit-Position: refs/heads/master@{#38191}
Drive-by fix: actually match the hint in the IsSpeculativeBinopMatcher.
Review-Url: https://codereview.chromium.org/2191883002
Cr-Commit-Position: refs/heads/master@{#38176}
This required the introduction of the CheckedNumberOrOddballAsWord32 use info, and a change in the RepresentationChanger to handle it.
BUG=
Review-Url: https://codereview.chromium.org/2184513003
Cr-Commit-Position: refs/heads/master@{#38086}
After multiplying two integers we emit code like:
if (result == 0) {
if (OR_OPERATION(rhs, lhs) < 0) {
DEOPT;
}
}
This CL allows us to eliminate the OR and comparison if either rhs or
lhs is a negative number, reducing the code to:
if (result == 0) DEOPT;
BUG=
Review-Url: https://codereview.chromium.org/2167643002
Cr-Commit-Position: refs/heads/master@{#38016}
Use the ForInFilterStub directly. Hence we will only jump to the runtime for
special receivers (instance_type <= LAST_SPECIAL_RECEIVER_TYPE) and for
converting element indices which are not in the string cache.
BUG=
Review-Url: https://codereview.chromium.org/2151773002
Cr-Commit-Position: refs/heads/master@{#37934}
This makes sure that we preserve call's tailness even if we have
introduced a loop exit between the call and the return.
BUG=chromium:628773
Review-Url: https://codereview.chromium.org/2155123002
Cr-Commit-Position: refs/heads/master@{#37832}
In int32 multiplication, if we have a positive integer as input, then we know we can't produce a -0 answer. The same is true if truncation is applied (x * y | 0). Without this information, we have to rather annoyingly check if the result of multiplication is 0, then OR the inputs to check for negativity, and possibly return -0. In TurboFan, we'll deopt in this case.
BUG=
Review-Url: https://codereview.chromium.org/2154073002
Cr-Commit-Position: refs/heads/master@{#37831}
Delaying for merges caused branch cloning using dummy effect phi inputs,
potentially splitting the effect chain at start.
We still have to delay the creation for loops because we need to break
cycles.
BUG=chromium:628403
Review-Url: https://codereview.chromium.org/2159603002
Cr-Commit-Position: refs/heads/master@{#37808}
This makes sure that the uses of PlainPrimitiveToNumber get a more
precise type (so that the uses know how to interpret the output
representation).
BUG=chromium:628516
Review-Url: https://codereview.chromium.org/2151223002
Cr-Commit-Position: refs/heads/master@{#37792}
Typed lowering now produces SpeculativeNumberShiftLeft for JSShiftLeft if the type feedback is kSignedSmall or kSigned32.
BUG=v8:4583
LOG=n
Review-Url: https://codereview.chromium.org/2150553002
Cr-Commit-Position: refs/heads/master@{#37762}
This drops the %_ValueOf intrinsic, but keeps the runtime entry
%ValueOf for now, by either migrating the functionality (mostly
Debug mirror or toString/valueOf methods) to C++ or TurboFan
builtins, or switching to the %ValueOf runtime call when it's
not performance critical anyways.
The %_ValueOf intrinsic was one of the last blockers for fixing
the unsound machine operator typing in TurboFan.
R=yangguo@chromium.org
BUG=v8:5049
Committed: https://crrev.com/293bd7882987f00e465710ce468bfb1eaa7d3fa2
Review-Url: https://codereview.chromium.org/2126453002
Cr-Original-Commit-Position: refs/heads/master@{#37512}
Cr-Commit-Position: refs/heads/master@{#37519}
Reason for revert:
[Sheriff] Breaks without i18n:
https://build.chromium.org/p/client.v8/builders/V8%20Linux%20-%20noi18n%20-%20debug/builds/8466
Original issue's description:
> [intrinsic] Drop the %_ValueOf intrinsic.
>
> This drops the %_ValueOf intrinsic, but keeps the runtime entry
> %ValueOf for now, by either migrating the functionality (mostly
> Debug mirror or toString/valueOf methods) to C++ or TurboFan
> builtins, or switching to the %ValueOf runtime call when it's
> not performance critical anyways.
>
> The %_ValueOf intrinsic was one of the last blockers for fixing
> the unsound machine operator typing in TurboFan.
>
> R=yangguo@chromium.org
> BUG=v8:5049
>
> Committed: https://crrev.com/293bd7882987f00e465710ce468bfb1eaa7d3fa2
> Cr-Commit-Position: refs/heads/master@{#37512}
TBR=yangguo@chromium.org,bmeurer@chromium.org
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=v8:5049
Review-Url: https://codereview.chromium.org/2117273002
Cr-Commit-Position: refs/heads/master@{#37514}
This drops the %_ValueOf intrinsic, but keeps the runtime entry
%ValueOf for now, by either migrating the functionality (mostly
Debug mirror or toString/valueOf methods) to C++ or TurboFan
builtins, or switching to the %ValueOf runtime call when it's
not performance critical anyways.
The %_ValueOf intrinsic was one of the last blockers for fixing
the unsound machine operator typing in TurboFan.
R=yangguo@chromium.org
BUG=v8:5049
Review-Url: https://codereview.chromium.org/2126453002
Cr-Commit-Position: refs/heads/master@{#37512}
The re-typer now only types a node if its inputs are all typed with the
exception of phi nodes. This works because all cycles in the graph have
to contain a phi node.
BUG=chromium:625558
Review-Url: https://codereview.chromium.org/2120243002
Cr-Commit-Position: refs/heads/master@{#37493}
Import fdlibm versions of acos, acosh, asin and asinh, which are more
precise and produce the same result across platforms (we were using
libm versions for asin and acos so far, where both speed and precision
depended on the operating system so far). Introduce appropriate TurboFan
operators for these functions and use them both for inlining and for the
generic builtin.
Also migrate the Math.imul and Math.fround builtins to TurboFan builtins
to ensure that their behavior is always exactly the same as the inlined
TurboFan version (i.e. C++ truncation semantics for double to float
don't necessarily meet the JavaScript semantics).
For completeness, also migrate Math.sign, which can even get some nice
love in TurboFan.
Drive-by-fix: Some alpha-sorting on the Math related functions, and
cleanup the list of Math intrinsics that we have to export via the
native context currently.
BUG=v8:3266,v8:3496,v8:3509,v8:3952,v8:5169,v8:5170,v8:5171,v8:5172
TBR=rossberg@chromium.orgR=franzih@chromium.org
Review-Url: https://codereview.chromium.org/2116753002
Cr-Commit-Position: refs/heads/master@{#37476}
The ARM64 instruction selector can generate code like this
negs w0, w1
b.vs deopt
but then reference the old value of w0 in the frame state, which will
obviously lead to wrong results.
R=jarin@chromium.org
BUG=v8:5158
Review-Url: https://codereview.chromium.org/2103793002
Cr-Commit-Position: refs/heads/master@{#37322}
This adds a dedicated test to make sure we don't try constant folding on
checks (in this case CheckTaggedPointer), which would generate invalid
code as we removing checks that guard the constant without knowing
whether it's safe to do so.
R=jarin@chromium.org
Review-Url: https://codereview.chromium.org/2087153002
Cr-Commit-Position: refs/heads/master@{#37168}
We cannot change x - y < 0 to x < y, because it would only be safe if
x - y cannot overflow, which we don't know in general.
R=jarin@chromium.org
BUG=v8:5129
Review-Url: https://codereview.chromium.org/2090493002
Cr-Commit-Position: refs/heads/master@{#37164}
This CheckBounds simplified operator is similar to the HBoundsCheck in
Crankshaft, and is hooked up to the new type feedback support in the
SimplifiedLowering. We use it to check the index bounds for keyed
property accesses.
Note to perf sheriffs: This will tank quite a few benchmarks, as the
operator makes some redundant branch elimination ineffective for
certain patterns of keyed accesses. This does require more serious
redundancy elimination, which we will do in a separate CL. So ignore
any regressions from this CL, we know there will be a few.
R=jarin@chromium.org
BUG=v8:4470,v8:5100
Committed: https://crrev.com/85e5567dae66a918500ae94c5568221137a0f5d4
Review-Url: https://codereview.chromium.org/2035893004
Cr-Original-Commit-Position: refs/heads/master@{#36947}
Cr-Commit-Position: refs/heads/master@{#37003}
This introduces SilenceNaN operator, which makes sure that we only
store quiet NaNs into holey arrays. We omit the NaN silencing code
at instruction selection time if the input is an operation that
cannot possibly produce signalling NaNs.
BUG=
Review-Url: https://codereview.chromium.org/2060233002
Cr-Commit-Position: refs/heads/master@{#36950}
This CheckBounds simplified operator is similar to the HBoundsCheck in
Crankshaft, and is hooked up to the new type feedback support in the
SimplifiedLowering. We use it to check the index bounds for keyed
property accesses.
Note to perf sheriffs: This will tank quite a few benchmarks, as the
operator makes some redundant branch elimination ineffective for
certain patterns of keyed accesses. This does require more serious
redundancy elimination, which we will do in a separate CL. So ignore
any regressions from this CL, we know there will be a few.
R=jarin@chromium.org
BUG=v8:4470,v8:5100
Review-Url: https://codereview.chromium.org/2035893004
Cr-Commit-Position: refs/heads/master@{#36947}
In Crankshaft we don't know reliably know that an HAdd might not turn
into a string addition later (via deoptimization), so we cannot set the
HValue::kAllowUndefinedAsNaN flag on the HAdd instruction in those
cases. It doesn't seem to affect performance if we just remove the flag
completely from the HAdd instruction, so let's stick to that approach
for now.
R=jarin@chromium.org
BUG=v8:5074
Review-Url: https://codereview.chromium.org/2048643002
Cr-Commit-Position: refs/heads/master@{#36805}
This introduces optimized number operations based on type feedback.
Summary of changes:
1. Typed lowering produces SpeculativeNumberAdd/Subtract for JSAdd/Subtract if
there is suitable feedback. The speculative nodes are connected to both the
effect chain and the control chain and they retain the eager frame state.
2. Simplified lowering now executes in three phases:
a. Propagation phase computes truncations by traversing the graph from uses to
definitions until checkpoint is reached. It also records type-check decisions
for later typing phase, and computes representation.
b. The typing phase computes more precise types base on the speculative types (and recomputes
representation for affected nodes).
c. The lowering phase performs lowering and inserts representation changes and/or checks.
3. Effect-control linearization lowers the checks to machine graphs.
Notes:
- SimplifiedLowering will be refactored to have handling of each operation one place and
with clearer input/output protocol for each sub-phase. I would prefer to do this once
we have more operations implemented, and the pattern is clearer.
- The check operations (Checked<A>To<B>) should have some flags that would affect
the kind of truncations that they can handle. E.g., if we know that a node produces
a number, we can omit the oddball check in the CheckedTaggedToFloat64 lowering.
- In future, we want the typer to reuse the logic from OperationTyper.
BUG=v8:4583
LOG=n
Review-Url: https://codereview.chromium.org/1921563002
Cr-Commit-Position: refs/heads/master@{#36674}
This prevents the compiler from optimizing
f64-to-tagged(tagged-to-f64(x)) ==> x
for non-number x (such as undefined).
Review-Url: https://codereview.chromium.org/2027593002
Cr-Commit-Position: refs/heads/master@{#36613}
Previously we first created a temporary graph for the inlinee and then
copied over all the nodes to the actual graph. This however introduces
unnecessary complexity, and we can instead just create the inlinee
inside the target graph.
R=jarin@chromium.org
Review-Url: https://codereview.chromium.org/2006353003
Cr-Commit-Position: refs/heads/master@{#36508}
This promotes the escape analysis from an experimental feature to be a
fully supported feature. The main goal is to unleach ClusterFuzz on the
implementation so that we can stabilize it.
R=jarin@chromium.org
Review-Url: https://codereview.chromium.org/1989833002
Cr-Commit-Position: refs/heads/master@{#36324}
