> On Apr 28, 2016, at 05:35, Nicolas B. Pierron <[email protected]> > wrote: > > and we would have to detail all the hidden assumptions which are currently > present in all the phases
As a newcomer to IonMonkey: Yes, please! ;-) I wrote the initial version of LLVM’s machine code verifier which can be run between the individual phases of the low-level code generator (not the higher-level LLVM IR verifier). We had similar problems that the required invariants would vary through the pipeline, and just verifying the intersection set of invariants that must hold everywhere was not good enough. We used flags on the IR data structure to indicate which invariants were required. For example, the flag MRI->isSSA() indicates that the virtual registers must be in SSA form. When the code is taken out of SSA form, the responsible pass would call MRI->leaveSSA(), and from that point on the machine code verifier no longer requires virtual registers to have a single definition. There’s a number of flags like that, see the “Function state” section of http://www.llvm.org/docs/doxygen/html/MachineRegisterInfo_8h_source.html <http://www.llvm.org/docs/doxygen/html/MachineRegisterInfo_8h_source.html> Unfortunately, running the machine code verifier after every pass is quite slow, so it is enabled with a command-line switch. It is useful for debugging crashes and miscompilations, but it is important to have test cases in CI that enable it, otherwise the checks quickly bit rot. Thanks, /jakob _______________________________________________ dev-tech-js-engine-internals mailing list [email protected] https://lists.mozilla.org/listinfo/dev-tech-js-engine-internals
