TheDan64 / inkwell

It's a New Kind of Wrapper for Exposing LLVM (Safely)
https://thedan64.github.io/inkwell/
Apache License 2.0
2.22k stars 217 forks source link

Fix memory leak caused by LLVMVerifyModule function call #507

Open antonbaliasnikov opened 2 weeks ago

antonbaliasnikov commented 2 weeks ago

Description

Fix a memory leak in the verify function that calls the LLVMVerifyModule function of LLVM.

If we build inkwell with Address sanitizer enabled: RUSTFLAGS="-Z sanitizer=address" cargo +nightly build --target aarch64-unknown-linux-gnu -Zbuild-std

And try to use the the inkwell::module::Module::verify function, then sanitizer will report the following memory leak (if the function succeeds):

==31844==ERROR: LeakSanitizer: detected memory leaks

Direct leak of 2 byte(s) in 2 object(s) allocated from:
    #0 0x56420a9edfda in strdup /rustc/llvm/src/llvm-project/compiler-rt/lib/asan/asan_interceptors.cpp:570:3
    #1 0x564210600990 in LLVMVerifyModule /__w/era-compiler-tester/era-compiler-tester/target-llvm/build-final/../../llvm/llvm/lib/Analysis/Analysis.cpp:94:20
    #2 0x56420bfd2fef in inkwell::module::Module::verify::hc511d5bc97d2c81f /usr/local/cargo/git/checkouts/inkwell-ae33c16eecfbdc50/c0821d7/src/module.rs:736:29
...

This is caused by not freeing the memory allocated in the strdup() call in the LLVMVerifyModule LLVM function .

This fix calls LLVMDisposeMessage to explicitly call free() for the err_str pointer fixing the memory leak.

Related Issue

N/A

How This Has Been Tested

Manually by building with sanitizers including this fix and calling verify function.

Option\<Breaking Changes>

N/A

Checklist

TheDan64 commented 2 weeks ago

@antonbaliasnikov there are test failures that need to be addressed

Edit: May be unrelated

antonbaliasnikov commented 2 weeks ago

@antonbaliasnikov there are test failures that need to be addressed

Edit: May be unrelated

Hey @TheDan64 ,

I checked tests locally for a failed configuration of LLVM 4.0 on Ubuntu 20.04. All tests passed:

cargo test --release --features llvm4-0 --verbose
...
test result: ok. 296 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 11.67s

I'm attaching the full log for the reference.

inkwell-llvm-4.0-log.txt

This CI error (invalid memory reference) is odd (hardware related?) and looks not feasible to debug from my side, especially without repository access (I can't even restart workflows to check if it's sporadic or not).

Could you, please, check it with a test PR from the master? Let me know if I can help somehow.

TheDan64 commented 2 weeks ago

Not sure why, but master started failing, even though it passed before. Perhaps GH changed hardware or something under us which revealed existing UB