Re: [Tarantool-patches] [PATCH luajit 1/4] x86/x64: Don't fuse loads across table.clear.

[Sergey Bronnikov via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

[-- Attachment #1: Type: text/plain, Size: 2913 bytes --]

Hi, Sergey!

thanks for the patch! LGTM

On 10.01.2025 16:07, Sergey Kaplun wrote:
> From: Mike Pall <mike>
>
> Reported by Peter Cawley.
>
> (cherry picked from commit 45c88b7963de2969a9a656c03ba06ad995d7fd5f)
>
> Load fusing optimization takes into account only the presence of the
> corresponding stores, but not any calls that may affect the table
> content. This may lead to the incorrect stores if the fusing
> optimization occurs across the `table.clear()` call, leading to
> inconsistent behaviour between the JIT and the VM.
>
> This patch adds the corresponding check.
>
> Sergey Kaplun:
> * added the description and the test for the problem
>
> Part of tarantool/tarantool#10709
> ---
>   src/lj_asm_x86.h                              |  1 +
>   .../lj-1117-fuse-across-table-clear.test.lua  | 36 +++++++++++++++++++
>   2 files changed, 37 insertions(+)
>   create mode 100644 test/tarantool-tests/lj-1117-fuse-across-table-clear.test.lua
>
> diff --git a/src/lj_asm_x86.h b/src/lj_asm_x86.h
> index 86ce3937..f47c460a 100644
> --- a/src/lj_asm_x86.h
> +++ b/src/lj_asm_x86.h
> @@ -465,6 +465,7 @@ static Reg asm_fuseload(ASMState *as, IRRef ref, RegSet allow)
>         }
>       } else if (ir->o == IR_ALOAD || ir->o == IR_HLOAD || ir->o == IR_ULOAD) {
>         if (noconflict(as, ref, ir->o + IRDELTA_L2S, 0) &&
> +	  noconflict(as, ref, IR_CALLS, 0) &&  /* Don't cross table.clear. */
>   	  !(LJ_GC64 && irt_isaddr(ir->t))) {
>   	asm_fuseahuref(as, ir->op1, xallow);
>   	return RID_MRM;
> diff --git a/test/tarantool-tests/lj-1117-fuse-across-table-clear.test.lua b/test/tarantool-tests/lj-1117-fuse-across-table-clear.test.lua
> new file mode 100644
> index 00000000..2f7c91d1
> --- /dev/null
> +++ b/test/tarantool-tests/lj-1117-fuse-across-table-clear.test.lua
> @@ -0,0 +1,36 @@
> +local tap = require('tap')
> +-- Test file to demonstrate LuaJIT's incorrect fusion across
> +-- `table.clear()`.
> +-- See also:https://github.com/LuaJIT/LuaJIT/issues/1117.
> +local test = tap.test('lj-1117-fuse-across-table-clear'):skipcond({
> +  ['Test requires JIT enabled'] = not jit.status(),
> +})
> +
> +local ffi = require('ffi')
> +local table_clear = require('table.clear')
> +
> +test:plan(1)
> +
> +local tab = {0}
> +local alias_tab = tab
> +local result_tab = {}
> +
> +jit.opt.start('hotloop=1')
> +
> +for i = 1, 4 do
> +  -- Load to be fused.
> +  local value = tab[1]
> +  -- Clear the alias table to trick the code flow analysis.
> +  table_clear(alias_tab)
> +  -- Need this cast to trigger load fusion. See `asm_comp()` for
> +  -- the details. Before the patch, this fusion takes the
> +  -- incorrect address of the already cleared table, which leads
> +  -- to the failure of the check below.
> +  result_tab[i] = ffi.cast('int64_t', value)
> +  -- Revive the value.
> +  tab[1] = 0
> +end
> +
> +test:samevalues(result_tab, 'no fusion across table.clear')
> +
> +test:done(true)

[-- Attachment #2: Type: text/html, Size: 3387 bytes --]