Several MIPS regression tests fail with "decoder.cpp: Assertion `isa<BranchInst>(res.branchCall->getNextNode())' failed"

[s3rvac]

When running regression tests with the Debug build of RetDec on the current master (1da7146), several MIPS tests fail with the following error:

retdec-bin2llvmir: src/bin2llvmir/optimizations/decoder/decoder.cpp:1424:
    void retdec::bin2llvmir::Decoder::handleDelaySlotLikely():
    Assertion `isa<BranchInst>(res.branchCall->getNextNode())' failed.
retdec-decompiler.sh: line 1264:  6781 Aborted   "$BIN2LLVMIR" "${BIN2LLVMIR_PARAMS[@]}" -o "$OUT_BACKEND_BC"
Error: Decompilation to LLVM IR failed

Here is a list of failing tests:

bugs.error-no-instructions-were-decoded.TestMips (mips-elf-e8e3b30ff295c1eac0cc06e1c8b1dc5e -k)     [ FAIL ]  (0.96s)
bugs.gcc-function-detection.gzip.mips.Test (gzip-strip -k)                                          [ FAIL ]  (1.45s)
bugs.llvmir-syntax-obj-defined-with-type.Test (mips-elf-fd27a58c009f1700499804d6f8a35981)           [ FAIL ]  (0.92s)

Steps to reproduce

• Checkout the current master and build RetDec in the Debug mode.
• Run the following commands:

$ wget https://github.com/avast-tl/retdec-regression-tests/raw/ec7d65bc95741aff954ae195ee9b1d247557867e/bugs/gcc-function-detection/gzip/mips/gzip-strip
$ retdec-decompiler.sh gzip-strip

The decompilation fails with the above error. The decompilation is expected to succeed (see this regression test).

Notes

• This error only appears with the Debug build of RetDec (in the Release build, assertions are disabled).
• This error is also present in the v3.1 release.
• This error was introduced by 3764c16 (prior to this commit, the decompilation finishes successfully).

Configuration

• Commit: 1da7146 (current master)
• 64b Arch Linux, GCC 8.1.0, Debug build of RetDec

[MerovingianByte]

When it says "partial bin2llvmir refactorization", does it mean there is more refactorization on the way?

[PeterMatula]

Yes, so far, I was able to refactor only some bin2llvmir passes, some of the others are still pretty bad. I'm nowhere near satisfied with the codebase quality of this most crucial RetDec part.