[maglev] Fix ResumeGenerator reviving a dead SuspendGenerator

ResumeGenerator is semantically a successor of SuspendGenerator (for
reasoning about liveness), but operationally it's a successor of
SwitchOnGeneratorState. This means that the jump to ResumeGenerator will
always create a new basic block, even if the SuspendGenerator was dead.

This causes problems if we made other assumptions on liveness based on
the semantics; in particular, we assume that JumpLoop is dead if the
loop header is dead (thanks to loop irreducibility).
SwitchOnGeneratorState breaks irreducibility, and this manifests as the
JumpLoop being alive and trying to jump to a dead header.

Since this is a special case, and loops are otherwise irreducible, we
can also fix it with a special case; namely, MarkBytecodeDead now has a
special case for SuspendGenerator which manually advances the iterator
and kills the ResumeGenerator.

Bug: v8:7700
Change-Id: Ice162f061e7ba1dda7ceb4f6fe9234889655b417
Fixed: v8:13250
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3865556
Reviewed-by: Victor Gomes <victorgomes@chromium.org>
Commit-Queue: Victor Gomes <victorgomes@chromium.org>
Auto-Submit: Leszek Swirski <leszeks@chromium.org>
Cr-Commit-Position: refs/heads/main@{#82931}
This commit is contained in:
Leszek Swirski 2022-09-01 13:50:12 +02:00 committed by V8 LUCI CQ
parent ed8be53861
commit 2f41221fbc

View File

@ -20,6 +20,7 @@
#include "src/interpreter/bytecode-array-iterator.h"
#include "src/interpreter/bytecode-decoder.h"
#include "src/interpreter/bytecode-register.h"
#include "src/interpreter/bytecodes.h"
#include "src/interpreter/interpreter-intrinsics.h"
#include "src/maglev/maglev-graph-labeller.h"
#include "src/maglev/maglev-graph-printer.h"
@ -179,6 +180,29 @@ class MaglevGraphBuilder {
// Any other bytecode that doesn't return or throw will merge into the
// fallthrough.
MergeDeadIntoFrameState(iterator_.next_offset());
} else if (bytecode == interpreter::Bytecode::kSuspendGenerator) {
// Extra special case for SuspendGenerator, if the suspend is dead then
// the resume has to be dead too. However, the resume already has a merge
// state, with exactly one predecessor (the generator switch), so it will
// be revived along the standard path. This can cause havoc if e.g. the
// suspend/resume are inside a dead loop, because the JumpLoop can become
// live again.
//
// So, manually advance the iterator to the resume, go through the motions
// of processing the merge state, but immediately emit an abort (which
// also kills the resume).
//
// TODO(leszeks): Instead of emitting an Abort, we could shrink the
// generator switch, removing this resume as an option.
iterator_.Advance();
DCHECK_EQ(iterator_.current_bytecode(),
interpreter::Bytecode::kResumeGenerator);
int resume_offset = iterator_.current_offset();
DCHECK_EQ(NumPredecessors(resume_offset), 1);
ProcessMergePoint(resume_offset);
StartNewBlock(resume_offset);
BuildAbort(AbortReason::kInvalidParametersAndRegistersInGenerator);
return;
}
// TODO(leszeks): We could now continue iterating the bytecode