Backports for 1.12 #57444
Backports for 1.12 #57444
Conversation
…57421) When loading a pkgimage, the new bpart validation code needs to check if the export set of any using'd packages differs from what it would have been during precompile. This could e.g. happen if somebody (or Revise) eval'd a new `export` statement into a package that was `using`'d. However, this case is somewhat rare, so let's optimize it by keeping a bit in `Module` that keeps track of whether anything like that has happened and if not skipping the revalidation. This slightly improves pkgimage load time in the ordinary case. More optimizations to follow. (cherry picked from commit 39255d4)
This addresses post-commit review #57230 (comment). This change was left-over from before I decided to also change the type of the `source` argument (at which point `source.module` was unavailable in the function). This module was supposed to be the same, but it turns out that both the julia tests and several packages use this code manually and use different modules for the two places. Use the same one we used before (which is probably more correct anyway) to fix #57417 (cherry picked from commit 0c5372f)
) This applies the existing prohibition against introducing new bindings in a closed module to binding replacement as well (for the same reason - the change won't be persisted after reload). It is pretty hard to even reach this point, since `eval` into closed modules is already prohibited and there's no surface syntax for cross-module declaration, but it is technically reachable from lowered IR. Further, in the future we may make all of these builtins, which would make it easier. Thus, be consistent now and fully disallow binding replacement in closed modules during precompile. (cherry picked from commit 56aed62)
…57426) An upcoming optimization will skip most binding validation if no binding replacement has taken place in (sysimage, pkgimage) modules. However, as a special case, we would like to treat `Main` as a non-sysimage module because the addition of new bindings in `Main` is common and we would like this to not ruin the optimization. To make this legal, we have to prohibit `import`ing or `using` any `Main` bindings in pkgimages. I don't think anybody actually does this, particularly, since `Main` is not considered loading during precompile (so you have to use the main binding via (Core|Base|).Main), and I can't think of any good semantic reason to want to do this, but regardless, it does add additional restrictions to `using`/`import`, so I wanted to break it out into its own PR. (cherry picked from commit 726c816)
…dules (#57433) This implements the optimization proposed in #57426 by keeping track of whether any bindings were replaced in image modules (excluding `Main` as facilitated by #57426). In addition, we augment serialization to keep track of whether a method body contains any GlobalRefs that point to a loaded (system or package) image. If both of these flags are true, we can skip scanning the body of the method, since we know that we neither need to add any additional backedges nor were any of the referenced bindings invalidated. The performance impact on end-to-end load time is small, but measurable. Overall `@time using ModelingToolkit` consistently improves about 5% using this PR. However, I should note that using time is still about 40% slower than 1.11. This is not necessarily an Apples-to-Apples comparison as there were substantial other changes on 1.12 (as well as current load-time-tunings targeting older versions), but I wanted to put the number context. (cherry picked from commit f6e2b98)
I was playing with strengthening the semantics around #57290. However, the particular change I was trying turned out too breaking (I may try a weaker version of it). Still, there were a number of good changes found in two categories: 1. We should explicitly import types when defining constructors. This has been allowed for a long time, but we may want to consider removing it, especially given the new binding semantics which make it more confusing as in #57290. 2. There were a couple of places where I don't think it was intended for generic functions in question not to extend Base. (cherry picked from commit 778e079)
Other backends (in this case NVPTX) require that `invariant.load` metadata is maintained to generate non-coherent loads. Currently, we unconditionally strip that metadata from all loads, since our other uses of it may have become invalid. x-ref: llvm/llvm-project#112834 JuliaGPU/CUDA.jl#2531 --------- Co-authored-by: Gabriel Baraldi <[email protected]> (cherry picked from commit 29da86b)
This reverts a portion of commit 50833c8. This algorithm is not able to handle simple cases where there is any internal padding, such as the example of: ``` struct LotsBytes a::Int8 b::NTuple{256,Int} c::Int end ``` Unfortunately fixing it is a bit of a large project right now, so reverting now to fix correctness while working on that. Fixes #55513 (indirectly, by removing broken code) Maybe reopens #54109, although the latency issue it proposes to fix doesn't occur on master even with this revert (just the mediocre looking IR result output returns) (cherry picked from commit a65c2cf)
also add nsw/nuw flags whenever possible. (cherry picked from commit b9a8d46)
Fixes #56904. The associated PR (#55876) compiles a finally block, then compiles a renumbered version of it. This works if `compile` doesn't mutate its input, but in reality, lambda bodies were being `set!` when linearized. The "invalid syntax" error was a result of attempting to linearize a lambda twice. (cherry picked from commit 414aca2)
…al (#57447) The `_DECLARED` partition kind used to be considered `guard`, but we now consider it equivalent to an Any-typed `_GLOBAL` (but with weaker redefinition properties). That said, its `->restriction` is NULL, so add it to the list of bindings that should return `nothing` here (and thus `Any` from the bulitin) to fix #57446. (cherry picked from commit 0163991)
|
@nanosoldier |
Backported PRs:
importing orusingMain during incremental compilation #57426--sysimage-native-code=nocmdlineargs tests single-threaded #57445newmpointer after it has been invalidated #57471linearize#57416Need manual backport:
Non-merged PRs with backport label:
RationalfromAbstractIrrational#55853