From: Sergey Kaplun via Tarantool-patches <tarantool-patches@dev.tarantool.org> To: Maxim Kokryashkin <m.kokryashkin@tarantool.org>, Sergey Bronnikov <sergeyb@tarantool.org> Cc: tarantool-patches@dev.tarantool.org Subject: [Tarantool-patches] [PATCH luajit 2/2] Fix canonicalization of +-0.0 keys for IR_NEWREF. Date: Wed, 10 May 2023 15:34:59 +0300 [thread overview] Message-ID: <e28ab7bb5f821f7276cb6c15c68c9fd54e3dacda.1683720396.git.skaplun@tarantool.org> (raw) In-Reply-To: <cover.1683720396.git.skaplun@tarantool.org> From: Mike Pall <mike> Reported by Sergey Kaplun. (cherry picked from commit 96fc114a7a3be3fd2c227d5a0ac53aa50cfb85d1) This commit is a follow-up for the commit f067cf638cf8987ab3b6db372d609a5982e458b5 ("Fix narrowing of unary minus."). Since this commit -0 IR constant is stored as well as +0 constant on the trace. Since IR NEWREF keys don't canonicalizied for -0 opposed of IR HREFK, it may lead to inconsistencies during trace recording. In particular, since -0 and 0 are different IR constants, alias analysis declares that they can't be aliased during folding optimization. Therefore: 1) For the IR TNEW we have non-nil value to lookup from the table via HLOAD, when only nil lookup is expected due to alias analysis. 2) For the IR TDUP we have non-nil value to lookup from the table via HLOAD, but the template table has no 0 field initiated as far as -0 isn't folding to 0 during parsing (see `bcemit_unop()` in <src/lj_parse.c>). These cases lead to the assertion failures in `fwd_ahload()`. This patch adds the aforementioned canonicalization. Sergey Bronnikov: * reported the original issue for the TDUP IR Sergey Kaplun: * added the description and the test for the problem Part of tarantool/tarantool#8516 --- Side note: I don't mention the 981 issue by intend -- I don't want to bother Mike with force pushes:). I suppose Igor should add this line (if he wants) went this commit will be cherry-picked to our master branch (a.k.a. tarantool). src/lj_record.c | 2 + .../tarantool-tests/lj-981-folding-0.test.lua | 57 +++++++++++++++++++ 2 files changed, 59 insertions(+) create mode 100644 test/tarantool-tests/lj-981-folding-0.test.lua diff --git a/src/lj_record.c b/src/lj_record.c index 9e2e1d9e..cc44db8d 100644 --- a/src/lj_record.c +++ b/src/lj_record.c @@ -1474,6 +1474,8 @@ TRef lj_record_idx(jit_State *J, RecordIndex *ix) TRef key = ix->key; if (tref_isinteger(key)) /* NEWREF needs a TValue as a key. */ key = emitir(IRTN(IR_CONV), key, IRCONV_NUM_INT); + else if (tref_isnumber(key) && tref_isk(key) && tvismzero(&ix->keyv)) + key = lj_ir_knum_zero(J); /* Canonicalize -0.0 to +0.0. */ xref = emitir(IRT(IR_NEWREF, IRT_PGC), ix->tab, key); keybarrier = 0; /* NEWREF already takes care of the key barrier. */ #ifdef LUAJIT_ENABLE_TABLE_BUMP diff --git a/test/tarantool-tests/lj-981-folding-0.test.lua b/test/tarantool-tests/lj-981-folding-0.test.lua new file mode 100644 index 00000000..251da24d --- /dev/null +++ b/test/tarantool-tests/lj-981-folding-0.test.lua @@ -0,0 +1,57 @@ +local tap = require('tap') +local test = tap.test('lj-981-folding-0'):skipcond({ + ['Test requires JIT enabled'] = not jit.status(), + ['Disabled on *BSD due to #4819'] = jit.os == 'BSD', +}) + +-- Test file to demonstrate LuaJIT misbehaviour on load forwarding +-- for -0 IR constant as table index. +-- See also, https://github.com/LuaJIT/LuaJIT/issues/981. + +local jparse = require('utils.jit_parse') + +jit.opt.start('hotloop=1') + +test:plan(4) + +-- Reset traces. +jit.flush() + +jparse.start('i') +local result +local expected = 'result' +-- TNEW: +-- -0 isn't folded during parsing, so it will be set with KSHORT, +-- UNM bytecodes. See <src/lj_parse.c> and bytecode listing +-- for details. +-- Because of it, empty table is created via TNEW. +for _ = 1, 4 do + result = ({[-0] = expected})[0] +end + +local traces = jparse.finish() + +-- Test that there is no any assertion failure. +test:ok(result == expected, 'TNEW and -0 folding') +-- Test that there is no NEWREF -0 IR. +test:ok(not traces[1]:has_ir('NEWREF.*-0'), '-0 is canonized for TNEW tab') + +-- Reset traces. +jit.flush() + +jparse.start('i') +-- TDUP: +-- Now just add a constant field for the table to use TDUP with +-- template table instead TNEW before -0 is set. +for _ = 1, 4 do + result = ({[-0] = expected, [1] = 1})[0] +end + +traces = jparse.finish() + +-- Test that there is no any assertion failure. +test:ok(result == expected, 'TDUP and -0 folding') +-- Test that there is no NEWREF -0 IR. +test:ok(not traces[1]:has_ir('NEWREF.*-0'), '-0 is canonized for TDUP tab') + +os.exit(test:check() and 0 or 1) -- 2.34.1
next prev parent reply other threads:[~2023-05-10 12:40 UTC|newest] Thread overview: 22+ messages / expand[flat|nested] mbox.gz Atom feed top 2023-05-10 12:34 [Tarantool-patches] [PATCH luajit 0/2] " Sergey Kaplun via Tarantool-patches 2023-05-10 12:34 ` [Tarantool-patches] [PATCH luajit 1/2] test: add utility for parsing `jit.dump` Sergey Kaplun via Tarantool-patches 2023-05-15 11:11 ` Maxim Kokryashkin via Tarantool-patches 2023-05-15 12:00 ` Maxim Kokryashkin via Tarantool-patches 2023-05-21 7:47 ` Sergey Kaplun via Tarantool-patches 2023-05-21 7:39 ` Sergey Kaplun via Tarantool-patches 2023-05-22 7:04 ` Sergey Kaplun via Tarantool-patches 2023-05-29 13:55 ` Maxim Kokryashkin via Tarantool-patches 2023-05-16 10:55 ` Sergey Bronnikov via Tarantool-patches 2023-05-22 7:02 ` Sergey Kaplun via Tarantool-patches 2023-05-22 9:14 ` Sergey Kaplun via Tarantool-patches 2023-05-10 12:34 ` Sergey Kaplun via Tarantool-patches [this message] 2023-05-15 12:05 ` [Tarantool-patches] [PATCH luajit 2/2] Fix canonicalization of +-0.0 keys for IR_NEWREF Maxim Kokryashkin via Tarantool-patches 2023-05-20 15:03 ` Sergey Kaplun via Tarantool-patches 2023-05-16 12:17 ` Sergey Bronnikov via Tarantool-patches 2023-05-20 14:54 ` Sergey Kaplun via Tarantool-patches 2023-05-22 7:55 ` Sergey Bronnikov via Tarantool-patches 2023-06-27 13:28 ` [Tarantool-patches] [PATCH luajit 1/3] test: split utils.lua into several modules Igor Munkin via Tarantool-patches 2023-06-27 13:35 ` Igor Munkin via Tarantool-patches 2023-06-28 11:36 ` Sergey Kaplun via Tarantool-patches 2023-06-28 16:07 ` Igor Munkin via Tarantool-patches 2023-07-04 17:10 ` [Tarantool-patches] [PATCH luajit 0/2] Fix canonicalization of +-0.0 keys for IR_NEWREF 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=e28ab7bb5f821f7276cb6c15c68c9fd54e3dacda.1683720396.git.skaplun@tarantool.org \ --to=tarantool-patches@dev.tarantool.org \ --cc=m.kokryashkin@tarantool.org \ --cc=sergeyb@tarantool.org \ --cc=skaplun@tarantool.org \ --subject='Re: [Tarantool-patches] [PATCH luajit 2/2] Fix canonicalization of +-0.0 keys for IR_NEWREF.' \ /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