252308 2023-02-15 03:02:37 -0800 REGRESSION(256816@main): [CMake] Build not reproducible due to __TIMESTAMP__ macro in generated file 2023-08-10 14:22:42 -0700 1 1 1 Unclassified WebKit JavaScriptCore WebKit Nightly Build Unspecified Unspecified RESOLVED DUPLICATE 258517 https://bugs.webkit.org/show_bug.cgi?id=247622 P2 Normal --- 1 berto webkit-unassigned alex.kanavin bugs-noreply emw mcatanzaro oldest_to_newest 1933771 0 berto 2023-02-15 03:02:37 -0800 The JSCBytecodeCacheVersion.cpp file contains the __TIMESTAMP__ macro which is replaced by the timestamp of the source file. This macro was already being used in 2.38.x and earlier versions, but at least building always from the same source tarball would produce the same timestamp so it would not affect the reproducibility of the build. However this is now generated by Source/JavaScriptCore/CMakeLists.txt so the timestamp varies every time WebKit is compiled from the same sources, making the build non-reproducible. 1933791 1 mcatanzaro 2023-02-15 06:53:28 -0800 Indeed, __TIMESTAMP__ is only deterministic when the source file is not generated, because it's based on last modification of the source file. The new comment in JSCBytecodeCacheVersion.h is also wrong except on Apple ports. The solution here should be to do the same thing Apple does: cat JSCBuiltins.o with CachedTypes.o and then hash that. They also use build version, but I'm not sure why and that can probably be ignored. This would likely be a good newcomers bug as it can be done entirely from CMake with no need to touch the source code files. 1934000 2 emw 2023-02-15 16:25:21 -0800 (In reply to Michael Catanzaro from comment #1) > Indeed, __TIMESTAMP__ is only deterministic when the source file is not > generated, because it's based on last modification of the source file. The > new comment in JSCBytecodeCacheVersion.h is also wrong except on Apple ports. > > The solution here should be to do the same thing Apple does: cat > JSCBuiltins.o with CachedTypes.o and then hash that. > They also use build version, but I'm not sure why and that can probably be ignored. We did this for extra safety, in case our production build environment was using a cached JSCBuiltins.o and CachedTypes.o *and* there was some binary incompatibility elsewhere in JSC. > This would > likely be a good newcomers bug as it can be done entirely from CMake with no > need to touch the source code files. FWIW, I tried to do the same change in CMake, and got stuck because the generated ninja build doesn't support inter-dependencies between object files in the same target. But I think we could make CMake hash the modification time relatively easy and restore the old semantics. 1963665 3 mcatanzaro 2023-06-26 13:56:46 -0700 *** Bug 258517 has been marked as a duplicate of this bug. *** 1963673 4 alex.kanavin 2023-06-26 14:09:34 -0700 To at least restore the binary reproducibility I made a PR where mtime of the generated file is set to the one it is generated from using 'touch -r': https://github.com/WebKit/WebKit/pull/15293 1970579 5 mcatanzaro 2023-08-10 14:22:42 -0700 *** This bug has been marked as a duplicate of bug 258517 ***