[Tarantool-patches] [PATCH luajit] Detect inconsistent renames even in the presence of sunk values.

[Igor Munkin]

Sergey,

Thanks for your review!

On 01.08.21, Sergey Kaplun wrote:
> Hi, Igor!
> Thanks for the patch!
> 
> Is there no point to simplify the test -- we have 5 different traces,
> when really need the only one (with RENAME between two possible jump
> to fallback branches with restoration from snapshot)?

Well... It was a tough issue to provide a stable reproducer to Mike.
Reducing this one is a much more complex issue: we need to compile a
trace for a loop with a rename emitted between two guards with the same
exitno in a variant part and leave the compiled loop via the guard at
the first iteration before the emitted RENAME. Sounds more complex than
even the existing test, doesn't it?

> Feel free to ignore.

Ignoring.

> 
> Side note: also, we should test that sunk optimization still works,
> shouldn't we?

Emm, nothing in the patch affects sink optimization per se, so if it's
OK without the patch, it's still OK with it.

> Feel free to ignore.

Ignoring.

> 
> Otherwise, LGTM.

Added your tag:
| Reviewed-by: Sergey Kaplun <skaplun at tarantool.org>

> 
> On 24.07.21, Igor Munkin wrote:
> > From: Mike Pall <mike>
> > 
> > Reported by Igor Munkin.
> > 
> > (cherry picked from commit 33e3f4badfde8cd9c202cedd1f4ed9275bc92e7d)
> > 
> > Side exits with the same exitno use the same snapshot for restoring
> > guest stack values. This obliges all guards related to the particular
> > snapshot use the same RegSP mapping for the values to be restored at the
> > trace exit. RENAME emitted prior to the guard for the same snapshot
> > leads to the aforementioned invariant violation.
> 
> This sentence a little bit unclear to me:
> 
> 1) Leads how?

RENAME changes the effective register to be used prior to the particular
snapshot[1]. Hence if RENAME is emitted between the guards with the same
exitno, RegSP mapping is inconsistent for the former one.

> 2) Do you mean it in terms of recording (i. e. the reverse instrucions
> recording order) or not?

The trace is recorded in a direct order, but *assembled* in a reverse
order. I implies the latter one.

> 
> >                                                  The easy way to save
> 
> > the snapshot consistency is spilling the renamed IR reference, that is
> > done in scope of <asm_snap_checkrename>.
> > 
> > However, the previous <asm_snap_checkrename> implementation considers
> > only the IR references explicitly mentioned in the snapshot. E.g. if
> > there is a sunk[1] object to be restored at the trace exit, and the
> > renamed reference is a *STORE to that object, the spill slot is not
> > allocated. As a result an invalid value is stored while unsinking that
> > object at all corresponding side exits prior to the emitted renaming.
> > 
> > To handle also those IR references implicitly used in the snapshot, all
> > non-constant and non-sunk references are added to the Bloom filter (it's
> > worth to mention that two hash functions are used to reduce collisions
> > for the cases when the number of IR references emitted between two
> > different snapshots exceeds the filter size). New <asm_snap_checkrename>
> > implementation tests whether the renamed IR reference is in the filter
> > and forces a spill slot for it as a result.
> > 
> > [1]: http://wiki.luajit.org/Allocation-Sinking-Optimization
> > 
> > Igor Munkin:
> > * added the description and the test for the problem
> > 
> > Resolves tarantool/tarantool#5118
> > Follows up tarantool/tarantool#4252
> > 
> > Signed-off-by: Igor Munkin <imun at tarantool.org>
> > ---
> > 
> > Related issues:
> > * https://github.com/tarantool/tarantool/issues/5118
> > * https://github.com/tarantool/tarantool/issues/4252
> > * https://github.com/LuaJIT/LuaJIT/issues/584
> > Branch: https://github.com/tarantool/luajit/tree/imun/lj-584-bad-renames-for-sunk-values
> > CI: https://github.com/tarantool/tarantool/commit/b35e2ee
> > 
> >  src/lj_asm.c                                  | 25 ++++---
> >  ...j-584-bad-renames-for-sunk-values.test.lua | 69 +++++++++++++++++++
> >  2 files changed, 81 insertions(+), 13 deletions(-)
> >  create mode 100644 test/tarantool-tests/lj-584-bad-renames-for-sunk-values.test.lua
> > 

<snipped>

> > -- 
> > 2.25.0
> > 
> 
> -- 
> Best regards,
> Sergey Kaplun

[1]: http://wiki.luajit.org/SSA-IR-2.0#miscellaneous-ops

-- 
Best regards,
IM