> On Feb 3, 2016, at 12:23 PM, Xinliang David Li <davi...@google.com> wrote: > > On Tue, Feb 2, 2016 at 1:31 PM, Bob Wilson <bob.wil...@apple.com > <mailto:bob.wil...@apple.com>> wrote: >> >>> On Jan 22, 2016, at 1:43 PM, Sean Silva via cfe-commits >>> <cfe-commits@lists.llvm.org> wrote: >>> >>> silvas added a comment. >>> >>> In http://reviews.llvm.org/D15829#333902, @davidxl wrote: >>> >>>> For the longer term, one possible solution is to make FE based >>>> instrumentation only used for coverage testing which can be turned on >>>> with -fcoverage-mapping option (currently, -fcoverage-mapping can not >>>> be used alone and must be used together with >>>> -fprofile-instr-generate). To summarize: >>>> >>>> A. Current behavior: >>>> >>>> ----------------------- >>>> >>>> 1. -fprofile-instr-generate turns on FE based instrumentation >>>> 2. -fprofile-instr-generate -fcoverage-mapping turns on FE based >>>> instrumentation and coverage mapping data generation. >>>> 3. -fprofile-instr-use=<..> assumes profile data from FE instrumentation. >>>> >>>> B. Proposed new behavior: >>>> >>>> -------------------------------- >>>> >>>> 1. -fprofile-instr-generate turns on IR late instrumentation >>>> 2. -fcoverage-mapping turns on FE instrumentation and coverage-mapping >>>> 3. -fprofile-instr-generate -fcoverage-mapping result in compiler warning >>>> 4. -fprofile-instr-use=<> will automatically determine how to use the >>>> profile data. >>> >>> >>> Very good observation that we can determine FE or IR automatically based on >>> the input profdata. That simplifies things. >>> >>>> B.2) above can be done today for improved usability. >>> >>> >>> I don't see how this improves usability. In fact, it is confusing because >>> it hijacks an existing option. >> >> Hijacking an existing option to do something different would definitely be a >> problem. Please find a way to specify IR-level instrumentation without >> breaking compatibility. If you want to replace the existing options with >> something different, we’ll need a transition period of at least 1-2 LLVM >> releases to migrate. >> > > If we remove B.3 above, then the proposed change (B.2) is essentially > making '-fcoverage-mapping' an alias to '-fprofile-instr-generate > -fcoverage-mapping'. No existing workflow will be broken and new > users can take advantage of the shortened option. The reason is that > there will be no users that only use -fcoverage-mapping option alone > and rely on its behavior (which is clang error).
The part I’m concerned about is B.1. The current behavior is that -fprofile-instr-generate enabled FE instrumentation. We can’t hijack that to do something different, at least without a sufficiently long transition period for clients to adapt. We use that to generate PGO profiles even when not using -fcoverage-mapping. > > >>> >>> Also B.3 causes existing user builds to emit a warning, which is annoying. >>> >>> I would propose the following modification of B: >>> >>> C.: >>> >>> 1. -fprofile-instr-generate defaults to IR instrumentation (i.e. behaves >>> exactly as before, except that it uses IR instrumentation) >>> 2. -fprofile-instr-generate -fcoverage-mapping turns on frontend >>> instrumentation and coverage. (i.e. behaves exactly as before) >>> 3. -fprofile-instr-use=<> automatically determines which method to use >>> >>> All existing user workflows continue to work, except for workflows that >>> attempt to `llvm-profdata merge` some old frontend profile data (e.g. they >>> have checked-in to version control and represents some special workload) >>> with the profile data from new binaries. >> >> The coverage mapping adds considerable cost. IR-level instrumentation has >> some problems that make it undesirable for our workflow, so we need a way to >> select front-end instrumentation without adding a bunch of unnecessary >> overhead (generating the coverage mapping when you’re not actually doing >> coverage testing). I disagree with your assessment that existing workflows >> would continue to “work” because ours would not. >> >>> >>> Concretely, imagine the following workflow: >>> >>> clang -fprofile-instr-generate foo.c -o foo >>> ./foo >>> llvm-profdata merge default.profraw -o new.profdata >>> llvm-profdata merge new.profdata >>> /versioncontrol/some-important-but-expensive-to-reproduce-workload.profdata >>> -o foo.profdata >>> clang -fprofile-instr-use=foo.profdata foo.c -o foo_pgo >>> >>> I think this is a reasonable breakage. We would need to add a note in the >>> release notes. Unfortunately this is not expected breakage if we claim to >>> have forward compatibility for profdata (which IIRC Apple requires; >>> @bogner?). >> >> Yes, that is a requirement for us. We need existing profdata to work with >> newer versions of clang (which is why IR-level instrumentation doesn’t work >> for us). >> > > profile-use can automatically detect FE based profile data and use it > properly. The question is whether we have a need to support merging > profiles from IR and FE instrumentation. I don’t think it makes sense to merge those. They seem like fundamentally different kinds of data. The “forward compatibility” requirement is about different versions of the FE-based profile data. > > > >>> But I think this case will be rare and exceptional enough that we can >>> tolerate it: >>> >>> - a simple immediate workaround is to specify `-fcoverage-mapping` (which >>> also adds some extra stuff, but works around the issue) >>> - Presumably >>> /versioncontrol/some-important-but-expensive-to-reproduce-workload.profdata >>> is regenerated with some frequency which is more frequent than upgrading >>> the compiler, and so it is likely reasonable to regenerate it alongside a >>> compiler upgrade, using the workaround until then. >> >> No, that assumption is not necessarily true for us. We need to be able to >> upgrade the compiler without breaking projects that we don’t control, and >> that includes regressing their performance because of an outdated profile. > > I want to understand how we can guarantee to support old (FE based) > profiles with new compilers. The region to counter id > mapping/assignment depends on how the AST is generated by the frontend > and how the AST is traversed. Do we have any guarantee that the new > compiler can generate them in the same order? How is that enforced? > The function structural hash generated may also be different (given > the same source). The FE-based instrumentation uses a custom traversal of the ASTs so that we can control the order and make sure it doesn’t change. It still depends on the way the ASTs are generated but the AST nodes that are relevant for this are unlikely to change in ways that would affect the instrumentation. I would love to have a better way to enforce that. The hashing scheme is specified in a way that does not tie it to the details of the compiler version. And, if we change the hashing, there is a design requirement that we maintain compatibility with the old hashing based on the version number in the profile data. That gives us a strong incentive to minimize changes to the hashing, but I’m not very happy with the current hashing, so I do hope we can introduce a new version soon. > > thanks, > > David > > > >> >>> >>> >>> >>>> B.1) needs a >>> >>>> transition period before the IR based instrumentation becomes >>> >>>> stablized (and can be flipped to the default). During the transition >>> >>>> period, the behavior of 1) does not change, but a cc1 option can be >>> >>>> used to turn on IR instrumentation (as proposed by Sean). >>> >>> >>> Just to clarify, users are not allowed to use cc1 options. The cc1 option >>> is purely for us as compiler developers to do integration and testing, put >>> off some discussions for later, etc. >>> >>> >>> http://reviews.llvm.org/D15829 >>> >>> >>> >>> _______________________________________________ >>> cfe-commits mailing list >>> cfe-commits@lists.llvm.org >>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
_______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
