## Summary <!-- Summarize the bug encountered concisely --> LLVM compiler generates wrong getelementptr instructions when accessing record containing record with variant parts. This simple rtl modification and test source was found after long analysis of the RTL heap errors using -gh option. ## System Information <!-- The more information are provided the easier it is to replicate the bug --> - **Operating system: Linux <!-- Windows, Linux (if possible, also name the distro), FreeBSD, Android, ... --> - **Processor architecture: tested on aarch64 (but probably not cpu specific) <!-- x86, x86-64, ARM, AARCH64, AVR, RISC-V, PowerPC, ... --> - **Compiler version: trunk <!-- 3.2, 3.2.2, 3.3, trunk, beta, ... (if possible, give also the git hash) --> - **Device:** <!-- Computer, Tablet, Mobile, Amiga, Microcontroller, ... --> ## Steps to reproduce <!-- How one can reproduce the issue - this is very important! --> Apply the following simple patch to trunk: ``` muller@gcc103:~/pas/trunk/fpcsrc$ git diff rtl/unix/cthreads.pp diff --git a/rtl/unix/cthreads.pp b/rtl/unix/cthreads.pp index 11dd9a00c1..84fa562e26 100644 --- a/rtl/unix/cthreads.pp +++ b/rtl/unix/cthreads.pp @@ -680,12 +680,20 @@ function IntBasicEventCreate(EventAttributes : Pointer; AManualReset,InitialStat {$if defined(Linux) and not defined(Android)} timespec: ttimespec; {$ifend} +{$ifdef DEBUG} + dist : ptrint; +{$endif} begin new(plocaleventstate(result)); plocaleventstate(result)^.FManualReset:=AManualReset; plocaleventstate(result)^.FWaiters:=0; plocaleventstate(result)^.FDestroying:=False; plocaleventstate(result)^.FIsSet:=InitialState; +{$ifdef debug} + dist:=ptrint(@plocaleventstate(result)^.FDestroying)-ptrint(result); + if dist > sizeof(plocaleventstate(result)^) then + writeln(stderr,'problem within IntBasicEventCreate'); +{$endif} {$if defined(Linux) and not defined(Android)} res := pthread_condattr_init(@plocaleventstate(result)^.FAttr); if (res <> 0) then ``` ## Example Project <!-- If possible, please create an example project that exhibits the problematic behavior, and link to it here in the bug report. --> Compile the following simple program ``` muller@gcc103:~/pas/check$ cat test-cthreads.pp {$mode objfpc} uses cthreads, classes; var i : longint; threadvar j : longint; begin i:=89; j:=i; writeln('j=',j); end. ``` Using LLVM compiler (recompiled with -dDEBUG to exposed the rtl modification above), using: ``` ppca64 -gwl test-cthreads.pp ./test-cthread ``` ## What is the current bug behavior? <!-- What actually happens --> ``` muller@gcc103:~/pas/check$ ./test-cthreads problem within IntBasicEventCreate j=89 An unhandled exception occurred at $0000000000431F64: EAccessViolation: Access violation $0000000000431F64 An unhandled exception occurred at $0000000000431F64: EAccessViolation: Access violation $0000000000431F64 $0000FFFEC4D07858 ``` ## What is the expected (correct) behavior? <!-- What you should see instead --> The output should only be: `` j=89 `` and no error should happen. ## Relevant logs and/or screenshots <!-- Paste any relevant logs - please use code blocks (```) to format console output, logs, and code, as it's very hard to read otherwise. You can also use syntax highlighting for Pascal with: ```pascal the code``` For more information see https://docs.gitlab.com/ee/user/markdown.html --> ## Possible fixes <!-- If you can, link to the line of code that might be responsible for the problem --> LLVM seems to have complicated strategies to cope with C-unions, and these probably need to be adapted correctly for Free Pascal.