[Tarantool-patches] [PATCH luajit] x86/x64: Fix math.ceil(-0.9) result sign.

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

From: Mike Pall <mike>

Reported by minoki.

(cherry-picked from commit 674afcd4e21d0cf64de3219d347557a0aed8ecc7)

The `ceil` (`floor`) math function implementation calculates (|x| +
2^52) - 2^52 for its argument to determine the fractional part of x , so
it will be rounded to the nearest integer and its sign is restored.
After that, if the original value is < (>) than the result, the -1 (1)
is subtracted from it. Take a look at the `ceil()` case. The result of
the operation `-1 - (-1)` is +0 for FP arithmetic, against -0 expected
as a result.

This patch changes the `- (-1)` operation to `+ 1` and restores sign
after it again.

NB: Since in DUALNUM mode on x86/x64 all results are tried to be
converted to integers the sign of 0 is neglected.

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

Part of tarantool/tarantool#9145
---

Branch: https://github.com/tarantool/luajit/tree/skaplun/lj-859-math-ceil-sign
Tarantool PR: https://github.com/tarantool/tarantool/pull/9326
Related issues:
* https://github.com/LuaJIT/LuaJIT/issues/859
* https://github.com/tarantool/tarantool/issues/9145

 src/vm_x64.dasc                               | 13 ++++++------
 src/vm_x86.dasc                               | 13 ++++++------
 .../lj-859-math-ceil-sign.test.lua            | 20 +++++++++++++++++++
 3 files changed, 32 insertions(+), 14 deletions(-)
 create mode 100644 test/tarantool-tests/lj-859-math-ceil-sign.test.lua

diff --git a/src/vm_x64.dasc b/src/vm_x64.dasc
index 399dfcbf..fc4c6259 100644
--- a/src/vm_x64.dasc
+++ b/src/vm_x64.dasc
@@ -400,9 +400,6 @@
 |.macro sseconst_1, reg, tmp		// Synthesize 1.0.
 |  sseconst_hi reg, tmp, 3ff00000
 |.endmacro
-|.macro sseconst_m1, reg, tmp		// Synthesize -1.0.
-|  sseconst_hi reg, tmp, bff00000
-|.endmacro
 |.macro sseconst_2p52, reg, tmp		// Synthesize 2^52.
 |  sseconst_hi reg, tmp, 43300000
 |.endmacro
@@ -2598,15 +2595,17 @@ static void build_subroutines(BuildCtx *ctx)
   |  addsd xmm1, xmm3			// (|x| + 2^52) - 2^52
   |  subsd xmm1, xmm3
   |  orpd xmm1, xmm2			// Merge sign bit back in.
+  |  sseconst_1 xmm3, RD
   |  .if mode == 1		// ceil(x)?
-  |    sseconst_m1 xmm2, RD		// Must subtract -1 to preserve -0.
   |    cmpsd xmm0, xmm1, 6		// x > result?
+  |    andpd xmm0, xmm3
+  |    addsd xmm1, xmm0			// If yes, add 1.
+  |    orpd xmm1, xmm2			// Merge sign bit back in (again).
   |  .else			// floor(x)?
-  |    sseconst_1 xmm2, RD
   |    cmpsd xmm0, xmm1, 1		// x < result?
+  |    andpd xmm0, xmm3
+  |    subsd xmm1, xmm0			// If yes, subtract 1.
   |  .endif
-  |  andpd xmm0, xmm2
-  |  subsd xmm1, xmm0			// If yes, subtract +-1.
   |.endif
   |  movaps xmm0, xmm1
   |1:
diff --git a/src/vm_x86.dasc b/src/vm_x86.dasc
index 9fa9a3f7..05a8267b 100644
--- a/src/vm_x86.dasc
+++ b/src/vm_x86.dasc
@@ -533,9 +533,6 @@
 |.macro sseconst_1, reg, tmp		// Synthesize 1.0.
 |  sseconst_hi reg, tmp, 3ff00000
 |.endmacro
-|.macro sseconst_m1, reg, tmp		// Synthesize -1.0.
-|  sseconst_hi reg, tmp, bff00000
-|.endmacro
 |.macro sseconst_2p52, reg, tmp		// Synthesize 2^52.
 |  sseconst_hi reg, tmp, 43300000
 |.endmacro
@@ -3089,15 +3086,17 @@ static void build_subroutines(BuildCtx *ctx)
   |  addsd xmm1, xmm3			// (|x| + 2^52) - 2^52
   |  subsd xmm1, xmm3
   |  orpd xmm1, xmm2			// Merge sign bit back in.
+  |  sseconst_1 xmm3, RDa
   |  .if mode == 1		// ceil(x)?
-  |    sseconst_m1 xmm2, RDa		// Must subtract -1 to preserve -0.
   |    cmpsd xmm0, xmm1, 6		// x > result?
+  |    andpd xmm0, xmm3
+  |    addsd xmm1, xmm0			// If yes, add 1.
+  |    orpd xmm1, xmm2			// Merge sign bit back in (again).
   |  .else			// floor(x)?
-  |    sseconst_1 xmm2, RDa
   |    cmpsd xmm0, xmm1, 1		// x < result?
+  |    andpd xmm0, xmm3
+  |    subsd xmm1, xmm0			// If yes, subtract 1.
   |  .endif
-  |  andpd xmm0, xmm2
-  |  subsd xmm1, xmm0			// If yes, subtract +-1.
   |.endif
   |  movaps xmm0, xmm1
   |1:
diff --git a/test/tarantool-tests/lj-859-math-ceil-sign.test.lua b/test/tarantool-tests/lj-859-math-ceil-sign.test.lua
new file mode 100644
index 00000000..df0ca329
--- /dev/null
+++ b/test/tarantool-tests/lj-859-math-ceil-sign.test.lua
@@ -0,0 +1,20 @@
+local tap = require('tap')
+
+-- Test file to demonstrate the incorrect LuaJIT's behaviour
+-- for `math.ceil(x)` when argument `x`: -1 < x < -0.5.
+-- See also https://github.com/LuaJIT/LuaJIT/issues/859.
+
+local test = tap.test('lj-859-math-ceil-sign')
+
+test:plan(1)
+
+local IS_DUALNUM = tostring(tonumber('-0')) ~= tostring(-0)
+local IS_X86_64 = jit.arch == 'x86' or jit.arch == 'x64'
+
+-- Use `tostirng()` to compare sign of the returned value.
+-- Take any value from the mentioned range. Chosen one is
+-- mentioned in the commit message.
+test:ok((IS_DUALNUM and IS_X86_64) or tostring(math.ceil(-0.9)) == '-0',
+        'correct zero sign')
+
+test:done(true)
-- 
2.42.0