[Tarantool-patches] [PATCH luajit] x86/x64: Don't sink IR_CONV num.int with -O-dce.
Sergey Kaplun
skaplun at tarantool.org
Wed Nov 18 17:22:42 MSK 2026
From: Mike Pall <mike>
Reported by aleksisch. Thanks to Sergey Kaplun.
(cherry picked from commit 592f9e3f549c7f176a9a7a0d70189559305f2eda)
The CONV num.int is sunk even with disabled DCE optimization (which is
required for sink optimization). Hence, the register allocated for CONV
IR is the same as for the next IR (ADD, for example). This leads to the
incorrect results when exiting to the interpreter since the register
content is modified by the next instruction.
This patch fixes this by disabling sinking of the `IR_CONV` with
disabled DCE.
Sergey Kaplun:
* added the description and the test for the problem
Part of tarantool/tarantool#13442
---
Branch: https://github.com/tarantool/luajit/tree/skaplun/lj-1537-conv-sink-without-dce
Related issues:
* https://github.com/LuaJIT/LuaJIT/issues/1537
* https://github.com/tarantool/tarantool/issues/13442
Side note: macOS (x86_64) CI is red due to runner problems.
src/lj_asm.c | 3 +-
.../lj-1537-conv-sink-without-dce.test.lua | 38 +++++++++++++++++++
2 files changed, 40 insertions(+), 1 deletion(-)
create mode 100644 test/tarantool-tests/lj-1537-conv-sink-without-dce.test.lua
diff --git a/src/lj_asm.c b/src/lj_asm.c
index f163b2e3..a4d6cf3f 100644
--- a/src/lj_asm.c
+++ b/src/lj_asm.c
@@ -933,7 +933,8 @@ static void asm_snap_alloc1(ASMState *as, IRRef ref)
}
} else {
RegSet allow;
- if (ir->o == IR_CONV && ir->op2 == IRCONV_NUM_INT) {
+ if (ir->o == IR_CONV && ir->op2 == IRCONV_NUM_INT &&
+ (as->flags & JIT_F_OPT_DCE)) {
IRIns *irc;
for (irc = IR(as->curins); irc > ir; irc--)
if ((irc->op1 == ref || irc->op2 == ref) &&
diff --git a/test/tarantool-tests/lj-1537-conv-sink-without-dce.test.lua b/test/tarantool-tests/lj-1537-conv-sink-without-dce.test.lua
new file mode 100644
index 00000000..4b332acb
--- /dev/null
+++ b/test/tarantool-tests/lj-1537-conv-sink-without-dce.test.lua
@@ -0,0 +1,38 @@
+local tap = require('tap')
+
+-- Test file to demonstrate LuaJIT incorrect sink of the `IR_CONV`
+-- with disabled DCE optimization.
+-- See also: https://github.com/LuaJIT/LuaJIT/issues/1537.
+local test = tap.test('lj-1537-conv-sink-without-dce'):skipcond({
+ ['Test requires JIT enabled'] = not jit.status(),
+})
+
+test:plan(1)
+
+jit.opt.start('hotloop=1', '-dce')
+
+-- Before the patch, the CONV num.int is sunk even with disabled
+-- DCE optimization (which is required for sink optimization).
+-- Hence, the register allocated for CONV IR is the same as for
+-- the next ADD IR, leading to the incorrect results when
+-- exiting to the interpreter.
+-- | 0004 xmm7 num CONV 0003 num.int
+-- | 0005 xmm7 > num SLOAD #4 T
+-- | 0006 xmm7 + num ADD 0005 +1
+-- | ...
+-- | 0009 ------------ LOOP ------------
+-- | 0010 xmm7 + num ADD 0006 +1
+-- | 0011 rbp + int ADD 0007 +1
+-- | .... SNAP #3 [ ---- ---- ---- 0004 0010 ]
+
+local t = {}
+local conv, n = 0, 0
+for _ = 1, 4 do
+ -- Need preserve the conversion, not int value in a snap.
+ conv = #t * 1
+ n = n + 1 -- Occupies the same register.
+end
+
+test:is(conv, 0, 'correct value is restored from snapshot')
+
+test:done(true)
--
2.56.0
More information about the Tarantool-patches
mailing list