Compiling with optimizations disabled results in runtime error

[matteo-meluzzi]

When running a genmc executable compiled with optimizations disabled, I get a runtime error.
When running genmc with the default compiler flags, everything runs smoothly.

How to reproduce:

1. checkout master branch
2. autoreconf -i to generate the configure script
3. ./configure AR=llvm-ar CXXFLAGS="-O0 -g" to generate Makefile
4. make genmc to generate genmc executable
5. ./genmc --print-exec-graphs tests/correct/litmus/LB2/variants/lb0.c
   outputs:

GenMC v0.10.0 (LLVM 15.0.7)
Copyright (C) 2023 MPI-SWS. All rights reserved.

*** Compilation complete.
*** Transformation complete.
Tip: Estimating state-space size. For better performance, you can use --disable-estimation.
libc++abi: terminating due to uncaught exception of type std::length_error: vector
zsh: abort      ./genmc --print-exec-graphs tests/correct/litmus/LB2/variants/lb0.c

In contrast the same steps except for
3) ./configure AR=llvm-ar
outputs

No errors were detected.
Number of complete executions explored: 3
Total wall-clock time: 0.11s

I am using MacOS 14, llvm@15 installed

[JonasOberhauser]

unlikely to be fixed. IIRC llvm requires at least -O2

[matteo-meluzzi]

I see. Thank you for the reply! Is there any online resources that explain why this is the case?

[JonasOberhauser]

Can't find anything now. Heck maybe I dreamed the whole thing up, or it has been fixed since Iooked at it 4 years ago?

But as I remember, llvm used some... uh... not so sound things that break under -O0.

Maybe @michaliskok can weigh in and provide some clarity

[JonasOberhauser]

Also just for the sake of it, try downgrading to LLVM 15.0.0

[michaliskok]

This took embarrassingly long but here's a possible explanation.

First, I could only reproduce this in macOS. I am not aware of LLVM having issues with -O0, but the crash anyway indicates that this is some issue in how internal genmc functions are handled.

The precise cause is still a bit vague to me, but the cause of the crashes seems to be that some function calls that are normally resolved internally got offloaded to some LLVM-provided utils.

If a patch like this fixes the crashes, I'll merge something similar in the next release.

diff --git a/src/Runtime/Interpreter.h b/src/Runtime/Interpreter.h
index c8d9b670..3694f0ab 100644
--- a/src/Runtime/Interpreter.h
+++ b/src/Runtime/Interpreter.h
@@ -585,11 +585,14 @@ public:
        void visitSelectInst(SelectInst &I);
 
        void visitCallInstWrapper(CallInstWrapper CIW);
-#if LLVM_VERSION_MAJOR < 11
-       void visitCallSite(CallSite CS) { visitCallInstWrapper(CallInstWrapper(CS)); }
-#else
+// #if LLVM_VERSION_MAJOR < 11
+//     void visitCallSite(CallSite CS) { visitCallInstWrapper(CallInstWrapper(CS)); }
+// #else
        void visitCallBase(CallBase &CB) { visitCallInstWrapper(CallInstWrapper(CB)); }
-#endif
+// #endif
+       void visitCallInst(CallInst &I) { visitCallInstWrapper(CallInstWrapper(I)); }
+       void visitInvokeInst(InvokeInst &I) { visitCallInstWrapper(CallInstWrapper(I)); }
+       void visitCallBrInst(CallBrInst &I) { visitCallInstWrapper(CallInstWrapper(I)); }