We implement loop peeling for wasm, currently available behind a flag.
Loops are peeled regardless of size.
Bug: v8:11510
Change-Id: Ia4c883abdee83df632b2611584d608c44e3295c8
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3367615
Reviewed-by: Andreas Haas <ahaas@chromium.org>
Reviewed-by: Tobias Tebbi <tebbi@chromium.org>
Commit-Queue: Manos Koukoutos <manoskouk@chromium.org>
Cr-Commit-Position: refs/heads/main@{#78496}
We switch the order of inlining and loop unrolling optimizations. This
gives small improvements to wasm-gc benchmarks.
Changes:
- Change the loop analysis algorithm to accept loops directly connected
to the graph's end. This is required because some nodes in an inlined
function, such as tail calls, might be directly connected to the outer
function's end without an intervening LoopExit node.
- Based on the above, skip emitting loop exits for some Throw nodes in
WasmGraphBuildingInterface.
- Introduce WasmInliningPhase, add it before loop unrolling. Remove
inlining from WasmOptimizationPhase.
- Handle graph terminators in loop unrolling.
- Add loops in the inlined function to the callers loop_infos.
Drive-by:
- Allow more wasm builtins in unrolled loops.
- Reduce inlining parameters to reflect that functions are now slightly
smaller during inlining, as no unrolling has taken place yet.
Bug: v8:12166
Change-Id: Iadd6b2f75170aa153ca1efb47fbb0d185c2b8371
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3329783
Reviewed-by: Jakob Kummerow <jkummerow@chromium.org>
Reviewed-by: Nico Hartmann <nicohartmann@chromium.org>
Commit-Queue: Manos Koukoutos <manoskouk@chromium.org>
Cr-Commit-Position: refs/heads/main@{#78394}
On 32-bit architectures, we need to run Int64Lowering on the inlinee
code to make it compatible with the caller code.
Since Int64Lowering now runs while a GraphReducer is active, only one of
them can use node marks to store node states. Therefore, we move the
Int64Lowering node states to an internal map.
Bug: v8:12166
Change-Id: I53b85442d503e71fa533e06568f4b9db572a4401
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3283072
Reviewed-by: Jakob Kummerow <jkummerow@chromium.org>
Reviewed-by: Andreas Haas <ahaas@chromium.org>
Commit-Queue: Manos Koukoutos <manoskouk@chromium.org>
Cr-Commit-Position: refs/heads/main@{#77941}
In each wasm CallDescriptor, we store the signature of the call based on
the real parameters passed to the call. This signature is more precise
than the formal function signature. We use this signature in inlining
to enable more optimizations.
Changes:
- Add wasm_sig_ field to CallDescriptor.
- Construct the real signature in {DoCall} and {DoReturnCall} in
graph-builder-interface, and pass it to all call-related functions in
WasmGraphBuilder.
- Update {ReplaceTypeInCallDescriptorWith} to use ValueType over
MachineType. Construct the updated function signature.
- In wasm-inlining, kill the Call node after inlining.
- Add two tests.
Bug: v8:11510
Change-Id: Ica711b6b4d83945ecb7201be26577eab7db3c060
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3270539
Reviewed-by: Jakob Kummerow <jkummerow@chromium.org>
Reviewed-by: Nico Hartmann <nicohartmann@chromium.org>
Commit-Queue: Manos Koukoutos <manoskouk@chromium.org>
Cr-Commit-Position: refs/heads/main@{#77889}
This CL improves wasm inlining heuristics in Turbofan, for an average
8,5% performance improvement in selected benchmarks.
Changes:
- In WasmInliner::Reduce(), only collect inlining candidates into a
priority queue, according to WasmInliner::LexicographicOrdering.
Move actual inlining to Finalize().
- Remove the InlineFirstFew heuristic. Add two limits to inlining:
Maximum relative size increase (reversely proportional to the function
size), and absolute size increase.
- Pass information about call frequency from liftoff-collected feedback
to the WasmInliner though the wasm module.
- Run wasm inlining along other optimizations in the pipeline.
- Split inlining and speculative inlining tests.
Bug: v8:7748, v8:12166
Change-Id: Iccee22093db765981889a24451fb458dfce1f1a6
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3222764
Reviewed-by: Nico Hartmann <nicohartmann@chromium.org>
Reviewed-by: Jakob Kummerow <jkummerow@chromium.org>
Commit-Queue: Manos Koukoutos <manoskouk@chromium.org>
Cr-Commit-Position: refs/heads/main@{#77428}
Loop exits are only used during loop unrolling and are then removed, as
they cannot be handled by later optimization stages. Since unrolling
comes before inlining in the compilation pipeline, we should not emit
loop exits in inlined functions.
Bug: v8:12166
Change-Id: I28b3ebaf67c9e15b127eeb1a63906c4ecfd77480
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3195871
Reviewed-by: Clemens Backes <clemensb@chromium.org>
Commit-Queue: Manos Koukoutos <manoskouk@chromium.org>
Cr-Commit-Position: refs/heads/main@{#77175}
Behind the --wasm-inlining flag, we introduce speculative direct calls
as an alternative to invoking functions through references.
In pseudocode, call_ref(func_ref, args...) reduces to
if (func_ref == function_reference_at(expected_index)) {
call_direct(expected_index, args...)
} else call_ref(func_ref, args...)
The introduced direct call can later get inlined in WasmInliningPhase.
Currently, we always speculate that the reference is the function at
index 0. Proper heuristics, based on liftoff runtime feedback, will come
later.
Bug: v8:12166, v8:7748
Change-Id: Icd1319d3091b436e71906717fd8a2662bfbb8481
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3162602
Commit-Queue: Manos Koukoutos <manoskouk@chromium.org>
Reviewed-by: Jakob Kummerow <jkummerow@chromium.org>
Cr-Commit-Position: refs/heads/main@{#76884}
We add an option to BuildTFGraph to not emit stack checks and call
tracing and use it in inlined functions.
Also, we add tests for zero/multiple return values, as well as infinite
loops in the inlined function.
Bug: v8:12166
Change-Id: I5f34c57d9870592085804853ff23ba94897cc8d5
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3141589
Reviewed-by: Jakob Kummerow <jkummerow@chromium.org>
Commit-Queue: Manos Koukoutos <manoskouk@chromium.org>
Cr-Commit-Position: refs/heads/main@{#76718}
We introduce basic wasm inlining infrastructure behind a flag. The
implementation is currently incomplete. Additionally, we always inline
the function at index 0; proper inlining heuristics will be added later.
Changes:
- Rename WasmInliningPhase -> JSWasmInliningPhase
- Introduce WasmInliningPhase and WasmInliner.
- Pass additional parameters as needed to GenerateCodeForWasmFunction.
- Remove EnsureEnd in WasmGraphAssembler. Create end node at the start
of compilation.
- Add a simple test.
Bug: v8:12166
Change-Id: Ifd7006ba378e9f74cd248b71e16869fbbb8a82be
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3141575
Commit-Queue: Manos Koukoutos <manoskouk@chromium.org>
Reviewed-by: Georg Neis <neis@chromium.org>
Cr-Commit-Position: refs/heads/main@{#76689}
