tarantool / luajit / 6624712844

88%

tarantool/master: 93%

push

github

IR_MIN/IR_MAX is non-commutative due to underlying FPU ops.

Thanks to Peter Cawley.

(cherry-picked from commit 7a2b83a0c)

Even after the commit c05d10330 ("Fix
math.min()/math.max() inconsistencies."), some of the corner cases (see
tests in the commit for details) for `math.min()`/`math.max()` still
inconsistent in the JIT and the VM. This happens because `IR_MIN` and
`IR_MAX` are marked as commutative, so their operands were swapped by
`asm_swapops()`. As a result, because `minsd`[1]/`maxsd`[2] instructions
don't change the sign byte of the destination register, its sign is
preserved if we compare 0 with -0. When we compare something with NaN,
the second source operand (either a NaN or a valid floating-point value)
is written to the result. Hence, swapping the operands changed the
resulting value.

This patch removes the commutative flag from the aforementioned IRs to
prevent swapping of their operands.

[1]: https://c9x.me/x86/html/file_module_x86_id_173.html
[2]: https://c9x.me/x86/html/file_module_x86_id_168.html

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

Part of tarantool/tarantool#9145

5337 of 5972 branches covered (0.0%)

Branch coverage included in aggregate %.

20484 of 23293 relevant lines covered (87.94%)

2752038.14 hits per line