Build systems, compilers, linkers, oh my!

TODOs

decouple thread0 allocation from jemalloc
- sanitizer integration
- malloc not coupled to csu (could have libjemalloc)
- ability to replace with debug malloc, secure malloc, etc.
kernel IFUNC support - armv8.1 atomic instructions
Investigate autofdo (pgo)
Empirical questions out of pgo
Collect profile data from (buildworld libraries? package build cluster?)
Incrementally commit CHERIbuild changes

Can we ship LLVM tools (opt, llvm-config etc.) by default?
- Concerns with ABI stability - not provided by LLVM
- Problem: we already have /usr/lib/clang/4.0.0/ with no explicit ABI disclaimers…
- Post-summit update: not a concern right now, all static libs except for libclang_rt.asan-x86_64.so but can consider it not part of the supported ABI.

Build LLVM + Clang + cheribsd from MacOS or Linux
Now also builds postgres etc.
Wrote cmake build infrastructure for awk, sed, rmdir, etc. (build tools) then invoke bmake
Managed to get a FreeBSD MIPS build w/ cheribsd
Problem with generator tools (e.g. csh) that are not done in their own phase - they build in the same directory at the same time as the main binary
Separate build tool to own Makefile?
Incrementally migrate changes into FreeBSD upstream?
- hardcoded tool paths in tool-building shell scripts

CheriBSD increasingly building with external toolchain
Plan to migrate platforms using GCC today to external toolchain
For 12.0 - either use Clang or require xtoolchain
Release engineering / reproducibility, need to store xtoolchain packages used for release
Future: may treat local toolchain as special case of external toolchain?

2/4 major sanitizers mostly work on FreeBSD (dsan, ubsan)
Question: should certain sanitizers be a matter of tier support?
C source-to-source transformation for dynamic validation
- supervisor: Stephen Kell (https://www.cl.cam.ac.uk/~srk31)
- student work (not) at: https://github.com/rptynan/

CADETS project includes work to build some/all of the base system via LLVM IR
use cases:
- instrumentation (for security, dynamic validation, coverage…)
- LTO (overlap with existing LTO efforts?)
- make callgraph

pgo clang takes a very long time
- build llvm
- run through test inputs
- build llvm again
Google's autofdo can generate profile data from perf data
Q: how portable is profile data?
- A: pretty portable!
- could we build profile data on e.g., the package builders for distribution in the source tree?

Current analysis is based on gcov.
Coverage is statement based, would like condition decision based coverage.
- Lack of opensource tools means no ability to run under travis, etc
- Hard to know if we're writing correct/sufficient tests without coverage
- Support for “IR everywhere” may help in writing new/better tools