staticdata: Set min validation world to require world #57381
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
When we have implicit binding edges from literal GlobalRefs, these edges imply a implicit minimum world age of jl_require_age (which is what all bindings loaded from pkgimages get their definition age set to). In #57318, the `cpuid_llvm` minimum world age got set to `1` rather than `jl_require_world`, so codegen would refuse to perform the llvmcall special codegen, since it couldn't guarantee that the binding would actually resolve to `llvmcall` in all worlds. Fix that by adjusting staticdata.jl to set the appropriate minworld. That said, there are a few complicating factors here: 1. In general, most of our code doesn't handle world ranges with more than one partition. But for example, codegen could have checked the current world age and implicitly partitioned the binding at codegen time (in practice just adding an appropraite error check). 2. The included test case uses a cached inference. However, in the original issue it appeared that inlining was also creating new references to this replaced binding, which should not have been permitted. I have not fully investigated this behavior yet and this might be another bug. 3. In the original issue, the specialization had its max_world terminated a few ages past jl_require_world. I do not understand this behavior yet. Still, fixes #57318.
vtjnash
pushed a commit
that referenced
this pull request
Feb 13, 2025
KristofferC
pushed a commit
that referenced
this pull request
Feb 14, 2025
When we have implicit binding edges from literal GlobalRefs, these edges imply a implicit minimum world age of jl_require_age (which is what all bindings loaded from pkgimages get their definition age set to). In #57318, the `cpuid_llvm` minimum world age got set to `1` rather than `jl_require_world`, so codegen would refuse to perform the llvmcall special codegen, since it couldn't guarantee that the binding would actually resolve to `llvmcall` in all worlds. Fix that by adjusting staticdata.jl to set the appropriate minworld. That said, there are a few complicating factors here: 1. In general, most of our code doesn't handle world ranges with more than one partition. But for example, codegen could have checked the current world age and implicitly partitioned the binding at codegen time (in practice just adding an appropraite error check). 2. The included test case uses a cached inference. However, in the original issue it appeared that inlining was also creating new references to this replaced binding, which should not have been permitted. I have not fully investigated this behavior yet and this might be another bug. 3. In the original issue, the specialization had its max_world terminated a few ages past jl_require_world. I do not understand this behavior yet. Still, fixes #57318. (cherry picked from commit c6805e2)
KristofferC
pushed a commit
that referenced
this pull request
Feb 14, 2025
KristofferC
added a commit
that referenced
this pull request
Feb 17, 2025
Backported PRs: - [x] #57346 <!-- lowering: Only try to define the method once --> - [x] #57341 <!-- bpart: When backdating replace the entire bpart chain --> - [x] #57381 <!-- staticdata: Set min validation world to require world --> - [x] #57357 <!-- Only implicitly `using` Base, not Core --> - [x] #57383 <!-- staticdata: Fix typo in recursive edge revalidation --> - [x] #57385 <!-- bpart: Move kind enum into its intended place --> - [x] #57275 <!-- Compiler: fix unsoundness of getfield_tfunc on Tuple Types --> - [x] #57378 <!-- print admonition for auto-import only once per module --> - [x] #57392 <!-- [LateLowerGCFrame] fix PlaceGCFrameReset for returns_twice --> - [x] #57388 <!-- Bump JuliaSyntax to v1.0.2 --> - [x] #57266 <!-- 🤖 [master] Bump the Statistics stdlib from d49c2bf to 77bd570 --> - [x] #57395 <!-- lowering: fix has_fcall computation --> - [x] #57204 <!-- Clarify mathematical definition of `gcd` --> - [x] #56794 <!-- Make `Pairs` public --> - [x] #57407 <!-- staticdata: corrected implementation of jl_collect_new_roots --> - [x] #57405 <!-- bpart: Also partition the export flag --> - [x] #57420 <!-- Compiler: Fix check for IRShow definedness --> - [x] #55875 <!-- fix `(-Inf)^-1` inconsistency --> - [x] #57317 <!-- internals: add _defaultctor function for defining ctors --> - [x] #57406 <!-- bpart: Ignore guard bindings for ambiguity purposes --> - [x] #49933 <!-- Allow for :foreigncall to transition to GC safe automatically -->
KristofferC
removed
the
backport 1.12
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
When we have implicit binding edges from literal GlobalRefs, these edges imply a implicit minimum world age of jl_require_age (which is what all bindings loaded from pkgimages get their definition age set to). In #57318, the
cpuid_llvmminimum world age got set to1rather thanjl_require_world, so codegen would refuse to perform the llvmcall special codegen, since it couldn't guarantee that the binding would actually resolve tollvmcallin all worlds. Fix that by adjusting staticdata.jl to set the appropriate minworld.That said, there are a few complicating factors here:
In general, most of our code doesn't handle world ranges with more than one partition. But for example, codegen could have checked the current world age and implicitly partitioned the binding at codegen time (in practice just adding an appropraite error check).
The included test case uses a cached inference. However, in the original issue it appeared that inlining was also creating new references to this replaced binding, which should not have been permitted. I have not fully investigated this behavior yet and this might be another bug.
In the original issue, the specialization had its max_world terminated a few ages past jl_require_world. I do not understand this behavior yet.
Still, fixes #57318.