LGTM, thanks!

Hi, Sergey!
Thanks for the review!
Please consider my answers below.

On 25.06.25, Sergey Bronnikov wrote:

Hi, Sergey,

thanks for the patch! Please see my comments below.

Sergey

On 6/12/25 12:36, Sergey Kaplun wrote:

From: Mike Pall <mike>

Reported by caohongqing.
Fix contributed by Peter Cawley.

(cherry picked from commit 8fbd576fb9414a5fa70dfa6069733d3416a78269)

`asm_hrefk()` uses the check for the offset for the corresponding node
structure. However, the target load is performed from its inner `key`
field with the offset 8. In the case of a huge table, it is possible
that the offset of the node (4095 * 8) is less than 4096 * 8 and can be
emitted via the corresponding instruction as an immediate offset, but
the offset of the `key` field is not. This leads to the corresponding
assertion failure in `emit_lso()`.

The issue [1] contains yet another fix in the same place [2]. We decided 
to backport the patch

separately. But please mention this in commit message.


1. https://github.com/LuaJIT/LuaJIT/issues/1026

2. 
https://github.com/LuaJIT/LuaJIT/commit/93ce12ee15abf28ef4cb24ae7e4b8a5b73d75c85

These issues are completely independent, IMO. I would rather not mention
it. Otherwise, by this logic, we should mention every problem related to
the HREFK here.

This patch fixes this behaviour by the correct check.

Sergey Kaplun:
* added the description and the test for the problem

Part of tarantool/tarantool#11278
---

Related issues:
*https://github.com/LuaJIT/LuaJIT/issues/1026
*https://github.com/tarantool/tarantool/issues/11278
Branch:https://github.com/tarantool/luajit/tree/skaplun/lj-1026-arm64-invalid-hrefk-offset-check

  src/lj_asm_arm64.h                            |  2 +-
  ...-arm64-invalid-hrefk-offset-check.test.lua | 48 +++++++++++++++++++
  2 files changed, 49 insertions(+), 1 deletion(-)
  create mode 100644 test/tarantool-tests/lj-1026-arm64-invalid-hrefk-offset-check.test.lua

diff --git a/src/lj_asm_arm64.h b/src/lj_asm_arm64.h
index 6c7b011f..a7f059a2 100644
--- a/src/lj_asm_arm64.h
+++ b/src/lj_asm_arm64.h
@@ -885,7 +885,7 @@ static void asm_hrefk(ASMState *as, IRIns *ir)

<snipped>

diff --git a/test/tarantool-tests/lj-1026-arm64-invalid-hrefk-offset-check.test.lua b/test/tarantool-tests/lj-1026-arm64-invalid-hrefk-offset-check.test.lua
new file mode 100644
index 00000000..de243814
--- /dev/null
+++ b/test/tarantool-tests/lj-1026-arm64-invalid-hrefk-offset-check.test.lua
@@ -0,0 +1,48 @@
+local tap = require('tap')
+
+-- Test file to demonstrate LuaJIT misbehaviour when assembling
+-- HREFK instruction on arm64 with the huge offset.
+-- See also:https://github.com/LuaJIT/LuaJIT/issues/1026.
+local test = tap.test('lj-1026-arm64-invalid-hrefk-offset-check'):skipcond({
+  ['Test requires JIT enabled'] = not jit.status(),

It is an ARM-specific patch, should we add a condition for ARM here?

It is a good question. It was once discussed, and we decided not to add
the skip condition to make other architectures more covered by tests too
(for example, we may check MIPS/PPC in the same test if we want to
support them).

+})
+
+test:plan(1)
+
+-- The assertion fails since in HREFK we are checking the offset
+-- from the hslots of the table of the Node structure itself

s/Node/`Node`/

Fixed. See the iterative patch below.

===================================================================
diff --git a/test/tarantool-tests/lj-1026-arm64-invalid-hrefk-offset-check.test.lua b/test/tarantool-tests/lj-1026-arm64-invalid-hrefk-offset-check.test.lua
index de243814..caa6291d 100644
--- a/test/tarantool-tests/lj-1026-arm64-invalid-hrefk-offset-check.test.lua
+++ b/test/tarantool-tests/lj-1026-arm64-invalid-hrefk-offset-check.test.lua
@@ -10,7 +10,7 @@ local test = tap.test('lj-1026-arm64-invalid-hrefk-offset-check'):skipcond({
 test:plan(1)
 
 -- The assertion fails since in HREFK we are checking the offset
--- from the hslots of the table of the Node structure itself
+-- from the hslots of the table of the `Node` structure itself
 -- instead of its inner field `key` (with additional 8 bytes).
 -- So to test this, we generate a big table with constant keys
 -- and compile a trace for each HREFK possible.
===================================================================

Branch is force-pushed.

+-- instead of its inner field `key` (with additional 8 bytes).
+-- So to test this, we generate a big table with constant keys
+-- and compile a trace for each HREFK possible.
+
+local big_tab = {}
+-- The map of the characters to generate constant string keys.
+-- The offset of the node should be 4096 * 8. It takes at least
+-- 1365 keys to hit this value. The maximum possible slots in the
+-- hash part is 2048, so to fill it with the maximum density (with
+-- the way below), we need 45 * 45 = 2025 keys.
+local chars = 'abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRS'
+for c1 inchars:gmatch('.') do
+  for c2 inchars:gmatch('.') do
+    big_tab[c1 .. c2] = 1
+  end
+end
+
+jit.opt.start('hotloop=1')
+
+-- Generate bunch of traces.
+for c1 inchars:gmatch('.') do
+  for c2 inchars:gmatch('.') do
+    loadstring([=[
+      local t = ...
+      for i = 1, 4 do
+        -- HREFK generation.
+        t[ ']=] .. c1 .. c2 .. [=[' ] = i
+      end
+    ]=])(big_tab)
+  end
+end
+
+test:ok(true, 'no assertion failed')

I would replace testcase description to something like "emitted assembly 
is correct".

Feel free to ignore.

It triggers the assertion in the first place, so ignoring.

+
+test:done(true)