From: Igor Munkin via Tarantool-patches <tarantool-patches@dev.tarantool.org> To: Sergey Ostanevich <sergos@tarantool.org> Cc: tarantool-patches@dev.tarantool.org Subject: Re: [Tarantool-patches] [PATCH luajit] Detect inconsistent renames even in the presence of sunk values. Date: Tue, 3 Aug 2021 23:28:23 +0300 [thread overview] Message-ID: <20210803202823.GD27855@tarantool.org> (raw) In-Reply-To: <D5561952-CEEF-49E5-A0A5-1A76B619AD57@tarantool.org> Sergos, Thanks for your review! On 27.07.21, Sergey Ostanevich wrote: > Hi! Thanks for the patch! > > Just a small nit to the test. I won’t comment Mike’s code :) The fix is much more clearer and simpler than the test for it... > > LGTM Added your tag: | Reviewed-by: Sergey Ostanevich <sergos@tarantool.org> > > Sergos > > > On 24 Jul 2021, at 20:23, Igor Munkin <imun@tarantool.org> 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. 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@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> > > diff --git a/test/tarantool-tests/lj-584-bad-renames-for-sunk-values.test.lua b/test/tarantool-tests/lj-584-bad-renames-for-sunk-values.test.lua > > new file mode 100644 > > index 00000000..8aad3438 > > --- /dev/null > > +++ b/test/tarantool-tests/lj-584-bad-renames-for-sunk-values.test.lua > > @@ -0,0 +1,69 @@ > > +local tap = require('tap') > > + > > +local test = tap.test('lj-584-bad-renames-for-sunk-values') > > +test:plan(1) > > + > > +-- Test file to demonstrate LuaJIT assembler misbehaviour. > > +-- For more info, proceed to the issues: > > +-- * https://github.com/LuaJIT/LuaJIT/issues/584 > > +-- * https://github.com/tarantool/tarantool/issues/4252 > > + > > +----- Related part of luafun.lua. -------------------------------- > > + > > +local iterator_mt = { > > + __call = function(self, param, state) return self.gen(param, state) end, > > +} > > + > > +local wrap = function(gen, param, state) > > + return setmetatable({ > > + gen = gen, > > + param = param, > > + state = state > > + }, iterator_mt), param, state > > +end > > + > > +-- These functions call each other to implement a flat iterator > > +-- over the several iterable objects. > > +local chain_gen_r1, chain_gen_r2 > > + > > +chain_gen_r2 = function(param, state, state_x, ...) > > + if state_x ~= nil then return { state[1], state_x }, ... end > > + local i = state[1] + 1 > > + if param[3 * i - 1] == nil then return nil end > > + return chain_gen_r1(param, { i, param[3 * i] }) > > +end > > + > > +chain_gen_r1 = function(param, state) > > + local i, state_x = state[1], state[2] > > + local gen_x, param_x = param[3 * i - 2], param[3 * i - 1] > > + return chain_gen_r2(param, state, gen_x(param_x, state_x)) > > +end > > + > > +local chain = function(...) > > + local param = { } > > + for i = 1, select('#', ...) do > > + -- Put gen, param, state into param table. > > + param[3 * i - 2], param[3 * i - 1], param[3 * i] > > + = wrap(ipairs(select(i, ...))) > > + end > > + return wrap(chain_gen_r1, param, { 1, param[3] }) > > +end > > + > > +----- Reproducer. ------------------------------------------------ > > + > > +jit.opt.start(3, 'hotloop=3') > > I don’t like both numbers here. opt_level is 3 by default - why bother setting it? > And the second one should be factored out as an argument for both opt.start and the > loop below? Oops, this is the only place, that I didn't clean up... Yes, <opt_level> is excess: it is an artefact of juggling with options for reproducing. Now it's quite clear that the issue relates to allocation sinking optimization, that requires all flags to be enabled. Regarding the <hotloop> value, I dropped a verbose comment. Hope it makes the situation clearer, diff is below: ================================================================================ diff --git a/test/tarantool-tests/lj-584-bad-renames-for-sunk-values.test.lua b/test/tarantool-tests/lj-584-bad-renames-for-sunk-values.test.lua index 8aad3438..f037c898 100644 --- a/test/tarantool-tests/lj-584-bad-renames-for-sunk-values.test.lua +++ b/test/tarantool-tests/lj-584-bad-renames-for-sunk-values.test.lua @@ -51,7 +51,35 @@ end ----- Reproducer. ------------------------------------------------ -jit.opt.start(3, 'hotloop=3') +-- XXX: Here one can find the rationale for the 'hotloop' value. +-- 1. The most inner while loop on the line 86 starts recording +-- for the third element (i.e. 'c') and successfully compiles +-- as TRACE 1. However, its execution stops, since type guard +-- for <gen_x> result value on line 39 is violated (nil is +-- returned from <ipairs_aux>) and trace execution is stopped. +-- 2. Next time TRACE 1 enters the field is iterating through the +-- second table given to <chain>. Its execution also stops at +-- the similar assertion but in the variant part this time. +-- 3. <wrap> function becomes hot enough while building new +-- <chain> iterator, and it is compiled as TRACE 2. +-- There are also other attempts, but all of them failed. +-- 4. Again, TRACE 1 reigns while iterating through the first +-- table given to <chain> and finishes at the same guard the +-- previous run does. Anyway, everything above is just an +-- auxiliary activity preparing the JIT environment for the +-- following result. +-- 5. Here we finally come: <chain_gen_r1> is finally ready to be +-- recorded. It successfully compiles as TRACE 3. However, the +-- boundary case is recorded, so the trace execution stops +-- since nil *is not* returned from <ipairs_aux> on the next +-- iteration. +-- +-- JIT fine tuning via 'hotloop' option allows to catch this +-- elusive case, we achieved in a last bullet. The reason, why +-- this case leads to a misbehaviour while restoring the guest +-- stack at the trace exit, is described in the following LuaJIT +-- issue: https://github.com/LuaJIT/LuaJIT/issues/584. +jit.opt.start('hotloop=3') xpcall(function() for _ = 1, 3 do ================================================================================ If you're OK with the comment, I'll proceed with the patch. > > > + > > +xpcall(function() > > + for _ = 1, 3 do > > + local gen_x, param_x, state_x = chain({ 'a', 'b', 'c' }, { 'q', 'w', 'e' }) > > + while true do > > + state_x = gen_x(param_x, state_x) > > + if state_x == nil then break end > > + end > > + end > > + test:ok('All emitted RENAMEs are fine') > > +end, function() > > + test:fail('Invalid Lua stack has been restored') > > +end) > > + > > +os.exit(test:check() and 0 or 1) > > -- > > 2.25.0 > > > -- Best regards, IM
next prev parent reply other threads:[~2021-08-03 20:52 UTC|newest] Thread overview: 7+ messages / expand[flat|nested] mbox.gz Atom feed top 2021-07-24 17:23 Igor Munkin via Tarantool-patches 2021-07-27 13:53 ` Sergey Ostanevich via Tarantool-patches 2021-08-03 20:28 ` Igor Munkin via Tarantool-patches [this message] 2021-08-04 12:49 ` Vitaliia Ioffe via Tarantool-patches 2021-08-01 16:14 ` Sergey Kaplun via Tarantool-patches 2021-08-02 13:34 ` Igor Munkin via Tarantool-patches 2021-08-04 15:07 ` Igor Munkin via Tarantool-patches
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=20210803202823.GD27855@tarantool.org \ --to=tarantool-patches@dev.tarantool.org \ --cc=imun@tarantool.org \ --cc=sergos@tarantool.org \ --subject='Re: [Tarantool-patches] [PATCH luajit] Detect inconsistent renames even in the presence of sunk values.' \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: link
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox