[pull] main from llvm:main - #1665
Merged
Merged
Conversation
Update LLVM's DWARF language tables to match the current [`DW_LANG_*`](https://dwarfstd.org/languages.html) and [`DW_LNAME_*`](https://dwarfstd.org/languages-v6.html) registries. Add the missing `DW_LANG_*` and `DW_LNAME_*` constants, update the corresponding language-version mappings, and expose the new `DW_LANG_*` codes through the LLVM C API.
…gment (#206817) Introduce a new `PseudoProbeLinker` in `dsymutil` that collects `__probes/__probe_descs` sections from the debug map objects and merges the sections and places the probes metadata under `Contents/Resources/Profiling/{pseudo_probes[-<arch>],pseudo_probe_descs[-<arch>]}` respectively.
The [DWARF language registry](https://dwarfstd.org/languages.html) defines the default array lower bound for `DW_LANG_Fortran18` as one, but LLVM records it as zero. Update the language table to match the registry.
) RISCVISelDAGToDAG's lowering of RISCVISD::LPAD_CALL / LPAD_CALL_INDIRECT to PseudoCALLLpadAlign / PseudoCALLIndirectLpadAlign (introduced in #177515) only copied the callee, lpad label, chain, and glue operands, dropping the argument-register and register-mask operands in between. Without the register-mask operand, the register allocator treats these calls as clobbering nothing but ra, so values live across the call are not spilled/reloaded even though the callee is free to clobber caller-saved registers. This caused a miscompile where a pointer held live across a call to getcontext() (a returns_twice function) was corrupted after the call returned, leading to a SIGSEGV in llvm-test-suite's siod test. Fix the operand copy to include the argument-register and register-mask operands, matching the pseudo-instruction operands of a regular PseudoCALL/PseudoCALLIndirect. Signed-off-by: Jerry Zhang Jian <jerry.zhangjian@sifive.com>
a04f87f merged with some debug print leftovers. This patch removes them.
getLoopDepth is itself an O(depth) parent walk.
…ass (#211174) Just move it from pipeline to isel pass, NFC.
PerThreadAllocator, only used by DWARFLinker, indexes a fixed getThreadCount()-sized array by `getThreadIndex()`, so it only works on ThreadPoolExecutor threads (and asserts elsewhere). It false-shares adjacent threads' bump pointers. Instead, create each thread's sub-allocator lazily on first use, keyed by a process-unique instance id in a thread-local cache, with the instance owning the heap-allocated sub-allocators. This drops the getThreadIndex()/getThreadCount() dependence, unblocking the caller participation change for `parallelFor`. mlir::ThreadLocalCache solves the same problem by keying a per-thread map on the instance pointer and reclaiming a thread's slot when an instance dies, but costs a map lookup and shared_ptr bookkeeping per allocation; instances here are few and short-lived, so an id-indexed vector is cheaper. The counter behind claimPerThreadAllocatorId is defined out of line in Allocator.cpp so Windows DLLs cannot duplicate it and alias two instances' sub-allocators. Aided by Claude Fable 5
Profiling sqlite on aarch64-O3 shows ~0.5% of compile-time is spent in the SMEAttrs constructor, mostly from the inliner's areInlineCompatible and getInlineCallPenalty hooks. SMEAttrs does 14 string attribute lookups. Invert this to instead walk the function attributes once and build the mask with a StringSwitch. Improves CTMark geomean -0.20%, sqlite -0.40%. https://llvm-compile-time-tracker.com/compare.php?from=4aa1590ad66a1bff9b3c74b4a2c7366473015b4c&to=37796f0f92101225e2fb5b6924acfdb5cba1b672&stat=instructions%3Au Assisted-by: codex
…#210713) - builtins with documentation - instructions - new subtarget feature only for gfx13 - GlobalISel RegBankLegalize rules - tests
…cross binops (#211115) As discussed on #210948 - the "Op(shuffle(V1, Mask), shuffle(V2, Mask)) -> shuffle(Op(V1, V2), Mask)" fold should only occur in InstCombine if the shuffle is "like for like" and not length changing. Leave the more general fold to an upcoming VectorCombine patch which can handle the cost:benefit analysis, including multiuse shuffles.
Sub-dword args are widened to an i32 load whose extra bits belong to a sibling kernarg or padding, so noundef on the original arg does not cover the full loaded value Only copy it when the load type matches the argument type, mirroring the existing range/nofpclass guards
…onBlocks` (#210950) Problem ------- Under -mmlir --wrap-unstructured-constructs-in-execute-region, branch lowering hit a null Evaluation::block for body statements inside an OpenMP loop. genMultiwayBranch tripped Bridge.cpp: Assertion `block && "missing multiway branch block"' failed and genFIR(SelectCaseStmt) tripped Bridge.cpp: Assertion `e->block && "missing CaseStmt block"' failed The wrap machinery decides an unstructured DO/IF is wrappable and stops propagating its isUnstructured flag to the enclosing OpenMPConstruct. The OMP construct's isUnstructured stays false, so the top-level createEmptyBlocks treats it as a structured directive with nested evaluations — it allocates only the OMP construct's first-nested block and does not recurse into the DO body. Then OMP loop lowering takes over the body via createEmptyRegionBlocks, which was written to *re-parent* pre-existing blocks into its region and silently skipped evals whose .block was still null. Body targets (e.g. label 17's ContinueStmt, or a SelectCase's CaseStmts) therefore had no block when the consumer read evalOfLabel(...).block, tripping the asserts. Solution -------- Restore the invariant that consumers of eval.block can trust it at lowering time. In createEmptyRegionBlocks (DirectivesCommon.h), when an eval has isNewBlock but no .block, create one in the current region alongside the existing re-parenting path. The OMP lowering path then finds pre-allocated blocks for the whole body — matching the state that existed before the wrap feature changed isUnstructured propagation — without any lazy-allocation logic at consumer sites. Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
https://alive2.llvm.org/ce/z/dCmoVn In a GEP comparison with the same base like %1 = gep i8, @base, i64 a %2 = gep i8, @base, i64 b %cmp = icmp ... %1, %2 When we know that the offsets cross the base's alignment boundary the same number of times, it means that either both of them will overflow, or none of them will. In these cases we can turn the comparison into offset comparison: %cmp = icmp ... a, b
…way (#205564) As mentioned in #205562, there are cases where the transformation stage in LoopInterchange stops partway through, and the output IR ends up partially modified rather than interchanged. Notably, the interchange is recognized as having succeeded internally even in such cases. Apparently this happens not to cause any miscompiles at the moment, but it is clearly dangerous. This patch removes the early exit in the transformation phase. We cannot simply remove it, which checks for the presence of a unique successor of the inner loop header, because there is a case where the header actually has multiple successors. To avoid that situation, this patch changes the code to call SplitBlock on the inner loop header unconditionally. The test changes fall into two categories: some are simply due to newly added redundant BBs. In the others, the expected interchanges are now applied as intended, whereas previously the transformation failed and the interchanges were not applied before this patch.
Prototyping a new minimal type-based approach to RegBankSelect (#199040) for compile-time purposes exposed various gaps in instruction selection when not using the existing RegBankSelect pass. These manifested as new fallbacks when compiling the IR dataset from [1]. This patch teaches instruction selection to handle extracts of scalar i8/i16/i32 into GPR directly. This will prevent new fallbacks being introduced when a new type-based RBS pass is added. The test cases are extracted from [1]. Assisted-by: codex [1] https://davemgreen.github.io/gisel.html
Drop the Scenario dataclass as it is now clear what we are testing for.
In C, an array name in an expression "decays" into a pointer to its first element. LLDB did not honor this: commands like `memory read my_array` did not work correctly, because an aggregate type has no scalar value, so trying to obtain one (ResolveValue/GetValueAsUnsigned) failed. This MR adds explicit array decay: for array-typed expressions, the address of the array object itself is used instead of its (non-existent) scalar value.
…197437) Previously, the LangRef was ambiguous about the sign of comparisons used to create the loop dependence masks. This resulted in the expansion not following the intended semantics for extreme inputs. For example, %ptrA = 0, %ptrB = UINT_MAX, should result in a (RAW) mask with all lanes active. However, previously we'd do ``(%elementSize * lane) < abs(%ptrB - %ptrA)``, which due to incorrectly using signed arithmetic would result in a mask with a single lane active as ``abs(%ptrB - %ptrA)`` resulted in 1, not `UINT_MAX`. In other words, ``abs(%ptrB - %ptrA)`` should be ``unsigned-absolute-difference(%ptrA, %ptrB)``. Follow up to #188248.
…ock result (#209796) Previously, when the result of a try-lock call is branched on more than once, the paths between the branches would disagree on whether the capability is held while remaining consistent at each branch. The analysis then gave a false positive warning at the intermediate join: mutex 'lock' is not held on every path through here Create getTerminatorTrylockCall() helper from getEdgeLockset(); if the terminator of a block branches on the result of a call to a try_acquire_capability-function (perhaps negated or stored in a local variable), this helper returns that call and its callee. Use this new helper in getTerminatorTrylockCaps(), which will return the capabilities acquired by a trylock; feed these capabilites to intersectAndWarn() during a branch join, in order to avoid false positives. Soundness is preserved because intersectAndWarn() still removes the capabilities from the join block's entry lockset, ensuring any guarded access inside the join block before the terminator is diagnosed. getEdgeLockset() then re-adds the capability on the success edge. This change resolves the false positive warning reported in [1]. Reported-by: https://lore.kernel.org/linux-fsdevel/20260714181057.667977-1-timday@thelustrecollective.com/ [1] Assisted-by: AI was used a bit to understand the code and to help with testing and validation
…10571) `replaceWithAdditionalYields` only appends loop-carried values, so preserve the loop's discardable attributes, matching `scf.for`. Assisted-by: OpenAI Codex gpt-5.6-sol max
…wering (#209701) **Summary** This regression was introduced by #196094, which added a special lowering path for reductions on a single array element, such as `a(2)`. The problem is that Flang also treated an array section like `a(2:96)` as if it were a single element. Because of this, the section was sent to a code path that only supports scalar elements. That path produced an array type that the reduction initialization code could not handle, so Flang reached a `TODO` and aborted with a “not yet implemented” error. **Fix** The fix is to use the special element path only when the expression has rank 0, which means it represents one single value. Array sections have rank greater than 0, so they should continue through the existing boxed-array path. Fixes : [209462](#209462) --------- Co-authored-by: Jay Satish Kumar Patel <kumarpat@pe31.hpc.amslabs.hpecorp.net>
) When PHI-translating a load address through a select of module-level constant expressions, the folded side becomes a global constant whose use-list spans multiple functions. Scanning that use-list for an available cast could return a cast from another function, which then made DominatorTree::dominates() query a block from a different function and trip an assertion. This patch restricts the search to casts in the current function, matching the existing guards on the GEP and add paths. Fixes #211034
Add header generation support for langinfo.h, including nl_item type definition, langinfo macros, and YAML header specification. * libc/include/llvm-libc-types/nl_item.h: Define nl_item type * libc/include/llvm-libc-macros/langinfo-macros.h: Define POSIX macros * libc/include/langinfo.yaml: Define langinfo.h interface * libc/include/langinfo.h.def: Header template * libc/config/linux/*/headers.txt: Enable langinfo header for Linux Assisted-by: Automated tooling, human reviewed.
Lowers unpredicated stores of scalable vectors that are two or four times the width of a legal SVE type to multi-vector operations when subregister liveness is enabled instead of splitting them. This matches the existing approach for multi-vector load lowering.
…1163) This restriction dates back to the initial implementation of CoroElide support (dce9b02), but the rationale for treating ramp functions specially seems unclear and lacks test coverage. This patch proposes that we drop the restriction and enable more optimization opportunities. This also fixed the repro in #148380, but the root cause remains unresolved.
The Phabricator committing-a-change anchor no longer resolves since Phabricator has been discontinued, so drop the reference and keep the GitHub Issues guidance for linking reviewed/closed bugs.
Adds codegen support for the block and system-scoped NVVM atomic builtins: add, and, or, xor, min, max, inc and dec. These are lowered to the corresponding CIR `cir.atomic.fetch` operations and subsequently lowered to LLVM `atomicrmw` instructions.
…11230) Correct Propery->Property and sprinkle a couple of extra comments.
This file contains hidden or 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
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
20 participants
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.
See Commits and Changes for more details.
Created by
pull[bot] (v2.0.0-alpha.4)
Can you help keep this open source service alive? 💖 Please sponsor : )