Tarantool development patches archive
 help / color / mirror / Atom feed
From: Sergey Kaplun via Tarantool-patches <tarantool-patches@dev.tarantool.org>
To: Sergey Bronnikov <sergeyb@tarantool.org>,
	Evgeniy Temirgaleev <e.temirgaleev@tarantool.org>
Cc: tarantool-patches@dev.tarantool.org
Subject: [Tarantool-patches] [PATCH luajit] ARM64: Don't fuse sign extensions into logical operands.
Date: Mon, 16 Nov 2026 15:05:33 +0300	[thread overview]
Message-ID: <20261116120533.2886125-1-skaplun@tarantool.org> (raw)

From: Mike Pall <mike>

Thanks to Peter Cawley.

(cherry picked from commit 90742d91c27d185b70d1b4a6343fb6b7c26002db)

Logical instructions [1] (`orr`, etc.) have no sign extension form
(unlike `ADD Xd, Xn, Wm, SXTW #k`). This makes the fusing sign extension
into them meaningless and leads to the incorrect result in the
JIT-compiled code.

This patch prevents the fusing of sign extension (`sxtw`) into logical
instructions.

[1]: https://cybersandeep.gitbook.io/arm64basicguide/chapter-3-basic-arm64-instructions#id-3.2.2-logical-instructions

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-1076-not-fuse-sign-ext-logical
Related issues:
* https://github.com/LuaJIT/LuaJIT/issues/1076
* https://github.com/tarantool/tarantool/issues/13442

 src/lj_asm_arm64.h                            | 11 ++-
 ...lj-1076-not-fuse-sign-ext-logical.test.lua | 70 +++++++++++++++++++
 2 files changed, 74 insertions(+), 7 deletions(-)
 create mode 100644 test/tarantool-tests/lj-1076-not-fuse-sign-ext-logical.test.lua

diff --git a/src/lj_asm_arm64.h b/src/lj_asm_arm64.h
index f731ab05..f080dd6e 100644
--- a/src/lj_asm_arm64.h
+++ b/src/lj_asm_arm64.h
@@ -208,16 +208,13 @@ static Reg asm_fuseahuref(ASMState *as, IRRef ref, int32_t *ofsp, RegSet allow,
 static uint32_t asm_fuseopm(ASMState *as, A64Ins ai, IRRef ref, RegSet allow)
 {
   IRIns *ir = IR(ref);
+  int logical = (ai & 0x1f000000) == 0x0a000000;
   if (ra_hasreg(ir->r)) {
     ra_noweak(as, ir->r);
     return A64F_M(ir->r);
   } else if (irref_isk(ref)) {
-    uint32_t m;
     int64_t k = get_k64val(as, ref);
-    if ((ai & 0x1f000000) == 0x0a000000)
-      m = emit_isk13(k, irt_is64(ir->t));
-    else
-      m = emit_isk12(k);
+    uint32_t m = logical ? emit_isk13(k, irt_is64(ir->t)) : emit_isk12(k);
     if (m)
       return m;
   } else if (mayfuse(as, ref)) {
@@ -229,7 +226,7 @@ static uint32_t asm_fuseopm(ASMState *as, A64Ins ai, IRRef ref, RegSet allow)
 		    (IR(ir->op2)->i & (irt_is64(ir->t) ? 63 : 31));
       IRIns *irl = IR(ir->op1);
       if (sh == A64SH_LSL &&
-	  irl->o == IR_CONV &&
+	  irl->o == IR_CONV && !logical &&
 	  irl->op2 == ((IRT_I64<<IRCONV_DSH)|IRT_INT|IRCONV_SEXT) &&
 	  shift <= 4 &&
 	  canfuse(as, irl)) {
@@ -239,7 +236,7 @@ static uint32_t asm_fuseopm(ASMState *as, A64Ins ai, IRRef ref, RegSet allow)
 	Reg m = ra_alloc1(as, ir->op1, allow);
 	return A64F_M(m) | A64F_SH(sh, shift);
       }
-    } else if (ir->o == IR_CONV &&
+    } else if (ir->o == IR_CONV && !logical &&
 	       ir->op2 == ((IRT_I64<<IRCONV_DSH)|IRT_INT|IRCONV_SEXT)) {
       Reg m = ra_alloc1(as, ir->op1, allow);
       return A64F_M(m) | A64F_EX(A64EX_SXTW);
diff --git a/test/tarantool-tests/lj-1076-not-fuse-sign-ext-logical.test.lua b/test/tarantool-tests/lj-1076-not-fuse-sign-ext-logical.test.lua
new file mode 100644
index 00000000..b199648d
--- /dev/null
+++ b/test/tarantool-tests/lj-1076-not-fuse-sign-ext-logical.test.lua
@@ -0,0 +1,70 @@
+local tap = require('tap')
+
+-- The test file to demonstrate incorrect LuaJIT's fusing of the
+-- sign extension instruction into logical instructions.
+-- See also: https://github.com/LuaJIT/LuaJIT/issues/1076.
+local test = tap.test('lj-1076-not-fuse-sign-ext-logical'):skipcond({
+  ['Test requires JIT enabled'] = not jit.status(),
+})
+
+test:plan(5)
+
+jit.opt.start('hotloop=1')
+
+local ffi = require('ffi')
+local cast = ffi.cast
+
+local bnot = bit.bnot
+
+local N = -5
+
+-- Test avoiding erroneous fusing of:
+-- |  sxtw  x1, w4
+-- |  mvn   x0, x1
+--
+-- to the:
+-- | mov   x0, x4, lsl #48
+
+local r = {}
+for i = 1, 4 do
+  r[i] = bnot(cast('int64_t', N))
+end
+
+test:is(r[1], 4LL, 'correct result for bnot')
+test:samevalues(r, 'avoid fusion for explicit CONV with sign extension')
+
+-- Test avoiding erroneous fusing of:
+-- | sxtw  x21, w23
+-- | orr   x20, x20, x21
+--
+-- to the:
+-- | orn   x20, x20, x23, lsl #48
+
+local function logical_shift_unfused(subtest, expected)
+  subtest:plan(2)
+  jit.flush()
+  jit.opt.start('hotloop=1')
+  -- Use code generation to avoid using non-constant variables
+  -- defining functions preventing folding and fusion.
+  local script = string.format([[
+    local bor = bit.bor
+    local arshift, lshift, rshift = bit.arshift, bit.lshift, bit.rshift
+    local N = -5
+    local zero = 0LL
+    local result = {}
+    for i = 1, 4 do
+      result[i] = bor(%s(N, 1), zero)
+    end
+    return result
+  ]], subtest.name)
+  local f = assert(loadstring(script, 'unfused ' .. subtest.name))
+  local result = f()
+  subtest:is(result[1], expected, 'correct result for ' .. subtest.name)
+  subtest:samevalues(result, 'avoid fusion sxtw and orr for ' .. subtest.name)
+end
+
+test:test('arshift', logical_shift_unfused, -3LL)
+test:test('lshift',  logical_shift_unfused, -10LL)
+test:test('rshift',  logical_shift_unfused, 0x7FFFFFFDLL)
+
+test:done(true)
-- 
2.56.0


             reply	other threads:[~2026-11-16 12:06 UTC|newest]

Thread overview: 5+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-16 12:05 Sergey Kaplun via Tarantool-patches [this message]
2026-11-17 13:25 ` Evgeniy Temirgaleev via Tarantool-patches
2026-11-18  8:31   ` Sergey Kaplun via Tarantool-patches
2026-11-18  9:42     ` Evgeniy Temirgaleev via Tarantool-patches
2026-11-18 11:39       ` Sergey Kaplun 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=20261116120533.2886125-1-skaplun@tarantool.org \
    --to=tarantool-patches@dev.tarantool.org \
    --cc=e.temirgaleev@tarantool.org \
    --cc=sergeyb@tarantool.org \
    --cc=skaplun@tarantool.org \
    --subject='Re: [Tarantool-patches] [PATCH luajit] ARM64: Don'\''t fuse sign extensions into logical operands.' \
    /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