Hi Eli,
Thanks for sharing your experience. I’d very much like to fix the problems you encountered.
I’ve started looking at the state of code coverage recently; we figured LLVM itself would be a good test to figure out how mature it is, so I gave it a shot.
You may already be aware of this, but for readers who are not, there is a public bot which produces coverage reports for llvm roughly twice a day. You can find it by visiting llvm.org and clicking on the “llvm-cov” link within the “Useful Links” box (in the “Dev. Resources” section). Coverage is gathered by running check-{llvm,clang,polly,lld} and the ‘nightly’ test suite.
My experience:
- You have to specify -DLLVM_USE_LINKER=gold (or maybe lld works; I didn’t try). If you link with binutils ld, the program will generate broken profile information. Apparently, the linked binary is missing the __llvm_prf_names section. This took me half a day to figure out. This issue isn’t documented anywhere, and the only error message I got was “Assertion `!Key.empty()’ failed.” from llvm-cov.
I expect llvm-cov to print out “Failed to load coverage: ” in this situation. There was some work done to tighten up error reporting in ProfileData and its clients in r270020. If your host toolchain does have these changes, please file a bug, and I’ll have it fixed.
I was not aware of the issue with the binutils linker. We do have some end-to-end, runtime tests in compiler-rt which use this linker, so this type of failure is surprising. I’ve CC’d David Li, who has some experience working with this linker, in case he has any insight about the issue.
If you are using a relatively up-to-date host toolchain, I’ll add a note to our docs suggesting that users use gold when compiling with coverage enabled.
- The generated binaries are big and slow. Comparing to a build without coverage, llc becomes 8x larger overall (text section becomes roughly 2x larger). And check-llvm-codegen-arm goes from 3 seconds to 250 seconds.
The binary size increase comes from coverage mapping data, counter increment instrumentation, and profiling metadata.
The coverage mapping section is highly compressible, but exploiting the compressibility has proven to be tricky. I filed: llvm.org/PR33499.
Coverage makes use of frontend-based instrumentation, which is much less efficient than the IR-based kind. If we can find a way to map counters inserted by IR PGO to AST nodes, we could improve the situation. I filed: llvm.org/PR33500.
We can reduce testing time by not instrumented basic tools like count, not, FileCheck etc. I filed: llvm.org/PR33501.
- The generated profile information takes up a lot of space: llc generates a 90MB profraw file.
I don’t have any ideas about how to fix this. You can decrease the space overhead for raw profiles by altering LLVM_PROFILE_MERGE_POOL_SIZE from 4 to a lower value.
- When prepare-code-coverage-artifact.py invokes llvm-profdata for the profiles generated by “make check”, it takes 50GB of memory to process about 1.5GB of profiles. Is it supposed to use that much?
By default, llvm-profdata uses hardware_concurrency() to determine the number of threads to use to merge profiles. You can change the default by passing -j/–num-threads to llvm-profdata. I’m open to changing the ‘prep’ script to use -j4 or something like that.
- Using prepare-code-coverage-artifact.py generates “warning: 229 functions have mismatched data”. I’m not sure what’s causing this… I guess it has something to do with merging the profile data for multiple binaries? The error message is not very helpful.
This is unexpected. I’ll try to reproduce this, and I’ll fix the diagnostic along the way. I filed: llvm.org/PR33502.
- The HTML output highlights the semicolon after a break or return statement in some switch statements in red. (For example, LowerADDC_ADDE_SUBC_SUBE in ARMISelLowering.cpp.) Not really important, but annoying.
I’m sure I’m sitting on a bug report about this already, but unfortunately haven’t had the time to get around to it.
- On the bright side, when it works, the generated coverage information is precise and easy to read.
Good to hear.
vedant
Disk space is cheap, but the I/O takes a long time. I guess it’s specifically bad for LLVM’s “make check”, maybe not so bad for other cases. Oh, so it’s using a couple gigabytes per thread multiplied by 24 cores? Okay, now I’m not so worried.