Magic, non referenced variable changes it's value

[Fibonacci]

I was coding today and something strange happened. Somehow at some point part of my string was changed although I wasn't changing it (so I thought...)

Long story short, the string was changed, but by different function, which I didn't pass the variable as pointer or "var s: string"

I quickly came to the conclusion, if function is "inline" then parameters variables are passed as reference/pointer, but not always!

I wrote this POC code to show it in action. The strangest thing is the TWO functions MUST be inlined for the "magic" to happen, why is that? Is this kind of behavior documented somewhere? Some FPC dev shall explain?

Request: I don't have Delphi and I'm curious, please compile this code and post here if the string was replaced to "magic".

Quick poll: did you know about this? I didn't.

Code: Pascal  [Select][+][-]

1. program magic;
2.  
3. procedure test(s: string); inline;
4. begin
5.   //this procedure actually does nothing
6.   //it just changes it's "s" variable (copy of?) to "magic"
7.   s := 'magic';
8. end;
9.  
10. procedure main; inline;
11. var
12.   some_string: string;
13. begin
14.   //set some string
15.   some_string := 'some value';
16.  
17.   //call function that "does nothing"
18.   test(some_string);
19.  
20.   //expected "some value", shows "magic"
21.   writeln(some_string);
22.  
23.   readln;
24. end;
25.  
26. begin
27.   main;
28. end.

It's not because it's string, it does the same with any type, lets check integer:

Code: Pascal  [Select][+][-]

1. program magic;
2.  
3. procedure test(x: integer); inline;
4. begin
5.   x := 1234;
6. end;
7.  
8. procedure main; inline;
9. var
10.   int_test: integer;
11. begin
12.   //set some integer
13.   int_test := 555;
14.  
15.   //call function that "does nothing"
16.   test(int_test);
17.  
18.   //expected "555", shows "1234"
19.   writeln(int_test);
20.  
21.   readln;
22. end;
23.  
24. begin
25.   main;
26. end.
27.  

Regards

[Fibonacci]

Compiled in Delphi, different result. Works as expected. So seems like a bug in FPC.

[Blaazen]

Report it on GitLab. BTW, I have also opened one ticket about inlining: https://gitlab.com/freepascal.org/fpc/source/-/issues/40158

[Red_prig]

I have seen something similar, so you have to be careful with inline procedures

[WooBean]

@Fibonacci
« on: May 25, 2023, 01:15:06 am »

Well, so we have problem when calling inlined procedure inside inlined procedure when this called one (procedure/function) can change parameter(s) declared as value parameter(s). This behaviour of FPC breaks the statement obout parameters: 
"When parameters are declared as value parameters, the procedure gets a copy of the parameters that the calling statement passes. Any modifications to these parameters are purely local to the procedure’s block, and do not propagate back to the calling block."

One more observation, documentation says that inlining is disabled by default and needs  to be activated by directive {$INLINE ON} or commandline switch [-Si]. This option can be set inside Lazarus configuration. Actually (for Lazarus 2.2.0/FPC 3.2.2) inlining is accessible without it (for mode objFPC or delphi and 64/32 target).

[Bart]

IIRC then this has already been reported.

Bart

[Fibonacci]

I created a new issue on GitLab with some more info, disassemly and asmlists

https://gitlab.com/freepascal.org/fpc/source/-/issues/40291

[Bart]

As I thought: it was already reported: https://gitlab.com/freepascal.org/fpc/source/-/issues/39908

Bart