[Tarantool-patches] [PATCH luajit 2/2] Fix canonicalization of +-0.0 keys for IR_NEWREF.

[Sergey Bronnikov]

Hello, Sergey!

Thanks for the patch! To be honest I'm not proficient enough

for reviewing such patches, but I don't have objections, only a couple 
of minor comments. See below.

To give more confidence with fix I have started fuzzer for luaL_loadbuffer

with applied patch (no crashes for a about 5 hours).


Probably it is worth to mention that PUC Rio Lua has the same behaviour 
when table has indices "-0" and "0":

[0] ~/sources/MRG/tarantool/third_party/luajit$ lua
Lua 5.2.4  Copyright (C) 1994-2015 Lua.org, PUC-Rio
 > a = {[0] = 1, [2] = 4, [-0] = 7}
 > a[0]
7
 >

I believe it is important because you will say that Lua semantics will 
not broken after your patch.

Sergey


On 5/10/23 15:34, Sergey Kaplun wrote:
> 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')

You changed global JIT settings,

it is a good habit to put everything back when test is finished.

> +
> +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()


To be honest I think that chosen tables in tests are not representative. 
I propose to take this one:


local expected = 1
local result
for _ = 1, 4 do
   result = ({[0] = 1, [-0] = 2})[0]
end

assert(result == 2)

This example clearly demonstrates that element with index "0" was 
superseded by element with index "-0".


> +
> +-- 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)