MDEV-38670 Unary minus on empty string returns -0

[3titoo]

Normalize -0.0 to +0.0 in Item_func_neg::real_op() to match binary subtraction behavior. Add a test case in type_newdecimal.

[CLAassistant]

All committers have signed the CLA.

[ParadoxV5]

To a human, this looks like something a compiler might consider no-op.

[ParadoxV5]

… but, checking on https://godbolt.org/, it’s not the case for either GCC -O3 or Clang -O3.

[gkodinov]

This is a preliminary review.
Thank you for working on this!
You have pinpointed the problem correctly. But it needs more work to arrive at a complete solution, adequate testing and a solid commit message.

[gkodinov]

This is a very good start. But it's definitely not all there is to it.
This will only cover the negation function which you happen to be using.

The problem is more wide-spread. It's everywhere a FP literal is used.
E.g. try "select 0e0 * -1e0".

I believe this is happening because MySQL does IEEE 754 arithmetic. And doesn't account for "the signed zero" (https://en.wikipedia.org/wiki/Signed_zero) phenomenon while doing it.

As a result many operations done on the real and double data types internally can result in a signed zero.

The question now is where do you want to stop with the signed zero.

Approach 1: You could just acknowledge its presence in the docs. I.e. fully accept it.

Approach 2: dispose of it in the producer. Audit all implementations of real_op() in Item subclasses and call a utility function for their return value to turn the signed zero into an unsigned zero when detected.

Approach 3: dispose of it in the consumers. Every place where real_op() is called call a utility function to turn the signed zero into an unsigned zero when detected.

Which one to take is a question for the final review of course.
But at this point (before we start wasting domain expert's time) I'd like to request that you have a good story that covers more than just negation. And the story needs to be recorded into either the jira or the commit message itself.

[gkodinov]

Why do you need this? Do you actually have a test that hits this? I would be extremely surprised if you do.

[gkodinov]

FWIW, these are not floating point literals. The tests with these are irrelevant.

On how to specify constants in MariaDB, see:

MariaDB [test]> create table t1 as select 0 as a, 0.0 as b, 0e0 as c;
Query OK, 1 row affected (0.042 sec)
Records: 1  Duplicates: 0  Warnings: 0

MariaDB [test]> show create table t1;
+-------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| Table | Create Table                                                                                                                                                        |
+-------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| t1    | CREATE TABLE `t1` (
  `a` int(1) NOT NULL,
  `b` decimal(2,1) NOT NULL,
  `c` double NOT NULL
) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_uca1400_ai_ci |
+-------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1 row in set (0.001 sec)

Please add relevant tests that cover all of your changes and nothing but.

[gkodinov]

func_math is failing on most of the build-bot platforms. Please fix this.

[gkodinov]

It is customary to end a test file with a version test end echo command. Please add one for the target version: 13 in your case I believe, unless the final reviewer requests otherwise.