Open
Conversation
…ks (llvm#189522) Add `utils::diagDeprecatedCheckAlias` so checks can detect whether they are running under a deprecated name without enabling the new names. This commit also comes with an example with `zircon` module. It is deprecated in 22 release but we didn't provide a note for it before.
…#191797) These #includes are only needed in the SimpleRemoteEPC.cpp implementation.
…vm#191622) They have never existed since the initial public checkin.
For primitive array elements, we would accidentally activate the element and then immediate de-activate the array root, which is wrong. Ignore the element from the beginning to the later check never even compares with the element.
It's fine if they are uninitialized.
…ax. (llvm#191799) Extend test coverage with dedicated epilogue vectorization tests for dead first-order recurrences and FMinMaxNum reductions. Add users to FORs in existing tests where the dead FORs appeared unintentional.
As requested by @erichkeane here: llvm#190329 (comment)
…pe builtins (llvm#190969) When promoting scalar arguments to vectors for builtins like `ldexp`, `pown`, and `rootn`, use the correct vector type matching the argument element type instead of always using the return type: these builtins take an integer argument but at the same time have floating point return type Fix `ldexp` test that does not pass spirv-val and add similar tests for `pown` and `rootn` related to llvm#190736
Pass GISelValueTracking* through isKnownNeverNaN and isKnownNeverSNaN so that the implementation can call computeKnownFPClass to derive NaN information from value tracking, rather than only looking at flags and direct constant definitions. Update all callers. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…lvm#190519) Most clients don't have a notion of "address" and pass arbitrary values (including `0` and `sizeof(void *)`) to `DataExtractor` constructors. This makes address-extraction methods dangerous to use. Those clients that do have a notion of address can use other methods like `getUnsigned()` to extract an address, or they can derive from `DataExtractor` and add convenience methods if extracting an address is routine. `DWARFDataExtractor` is an example, where the removed methods were actually moved. This does not remove `AddressSize` argument of `DataExtractor` constructors yet, but makes it unused and overloads constructors in preparation for their deletion. I'll be removing uses of the to-be-deleted constructors in follow-up patches.
Enables LC_UUID load commands to be added with the addLoadCommand method. This will be used in future MachOPlatform changes to add support for adding UUIDs to MachO JITDylibs.
…#175870) Replace the manual check in `verifyRemoved()` with `AssertingVH` instrumentation. For cases where the leader table becomes very large, this is a cheaper way to verify we don't have dangling entries in the leader table. For this change, we must implement a move constructor for `AssertingVH` so that we can keep the first entry as an inline-allocated node that will be handled correctly as the table grows.
Part of the work to remove trivial VP intrinsics from the RISC-V backend, see https://discourse.llvm.org/t/rfc-remove-codegen-support-for-trivial-vp-intrinsics-in-the-risc-v-backend/87999 This PR expands four intrinsics before codegen, but doesn't remove the codegen handling yet as both DAGCombiner and type legalization can create these nodes. vp.fneg and vp.fpext are expanded in lockstep with the fma/fmuladd intrinsics since some test cases for vfmacc etc. also use these intrinsics, and mixing dynamic and constant vls causes some of the more complex patterns to be missed. The fixed-length VP vfmacc, vfmsac, vfnmacc and vfnmsac tests also need to replace the EVL of the vp.merge/vp.select with an immediate otherwise the resulting vmerge.vvm can't be folded into them. This only happens for fixed vector intrinsics with no passthru, since we end up with a constant vl from the fixed vector and dynamic vl from the vp.merge that prevents folding. As far as I'm aware we don't emit fixed length vp.merges in practice, since we only emit vp.merge in the loop vectorizer, and we only use it with EVL tail folding which requires a scalable VF.
This patch fixes 2 problems in lldb-server argument parser: 1. Let's try to start lldb-server with incorrect arguments ``` ./lldb-server platform --listen *:1111--server ``` Current behavior * lldb-server run in gdbserver mode with port 1111 Expected behavior * fail, as `1111–server` is not a number 2. And try to start lldb-server with host:port specification without colon ``` ./lldb-server gdbserver 1111 ./test Launched './test' as process 186... lldb-server-local_build lldb-server: llvm-project/lldb/source/Host/common/TCPSocket.cpp:245: virtual Status lldb_private::TCPSocket::Listen(llvm::StringRef, int): Assertion `error.Fail()' failed. PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace. Stack dump: 0. Program arguments: ./lldb-server gdbserver 1111 ./test Stack dump without symbol names (ensure you have llvm-symbolizer in your PATH or set the environment var `LLVM_SYMBOLIZER_PATH` to point to it): 0 lldb-server 0x0000002ab86d0ca2 1 lldb-server 0x0000002ab86ced06 2 lldb-server 0x0000002ab86d1428 3 linux-vdso.so.1 0x0000003f8e7fd800 __vdso_rt_sigreturn + 0 4 libc.so.6 0x0000003f8e2b264a 5 libc.so.6 0x0000003f8e27b1ac gsignal + 18 6 libc.so.6 0x0000003f8e26c14c abort + 180 7 libc.so.6 0x0000003f8e2760cc 8 libc.so.6 0x0000003f8e27610e __assert_perror_fail + 0 9 lldb-server 0x0000002ab86eb628 10 lldb-server 0x0000002ab86f1010 11 lldb-server 0x0000002ab86eeee0 12 lldb-server 0x0000002ab86eee5c 13 lldb-server 0x0000002ab863ef3a 14 lldb-server 0x0000002ab864067c 15 lldb-server 0x0000002ab86438da 16 libc.so.6 0x0000003f8e26c476 17 libc.so.6 0x0000003f8e26c51e __libc_start_main + 116 18 lldb-server 0x0000002ab863ce64 Aborted ``` We expect to see an error instead of lldb-server crash in this case
…#191359) Both implementations are currently equivalent. This is likely a leftover from the past, when `llvm::Optional` existed.
Expose existing trylock internal operation to posix interface. POSIX.1-2024 only specifies the `EBUSY` error case. Assisted-by: Codex with gpt-5.4 default fast
Part of the work to remove trivial VP intrinsics from the RISC-V backend, see https://discourse.llvm.org/t/rfc-remove-codegen-support-for-trivial-vp-intrinsics-in-the-risc-v-backend/87999 This splits off 4 intrinsics from llvm#179622.
…191819) This change aims to make it easier for MachOPlatform clients to customize JITDylib MachO headers. At MachOPlatform construction time clients can now supply a MachOPlatform::HeaderOptionsBuilder. The supplied callback will be called by setupJITDylib to create the HeaderOptions for the JITDylib being set up. No testcase: Constructing a MachOPlatform instance requires the ORC runtime, which we can't require for LLVM unit or regression suite tests. We should look at testing this functionality in the new ORC runtime once it's ready.
For a loop-nest-generating construct this function returns the number of loops in the generated loop nest. A loop-nest-transformation construct can be thought of as replacing N nested loops with K nested loops, where N = GetAffectedNestDepthWithReason K = GetGeneratedNestDepthWithReason
This change improves the lifetime safety checker to detect when
constructor parameters escape to class fields and suggest appropriate
`[[clang::lifetimebound]]` annotations.
```cpp
struct A {
View v;
A(const MyObj& obj) : v(obj) {} // Now suggests [[clang::lifetimebound]]
};
```
…lvm#191834) …cess). 6dbf9d1 forward declared the MemoryAccess class in ExecutorProcessControl.h, breaking some examples that were depending on the transitive include. (See e.g. https://lab.llvm.org/buildbot/#/builders/80/builds/21875). This commit adds the missing #includes to the broken examples.
These are now listed in the asciidoc spec here https://github.com/riscv/riscv-p-spec I got some help on this from AI, but I reviewed it. Test cases were fully generated with AI.
…m#191818) Add GICv5 `ICH_PPI_HVIR{0,1}_EL2` system registers (Interrupt Controller PPI Hide Virtual Interrupt Registers). These registers are added because a hypervisor may want to only expose a subset of the PPIs to the virtual machine and hide the remaining PPIs. The only way the hypervisor can do this is by trapping all the PPI ICV registers which leads to additional code complexity and adds performance overhead especially for nested virtualization. These are documented here: https://developer.arm.com/documentation/111107/latest/AArch64-Registers/ICH-PPI-HVIR-n--EL2--Interrupt-Controller-PPI-Hide-Virtual-Interrupt-Registers
…follow LLVM conventions (llvm#191134) Follow-up to #[189948](llvm#189948 (comment)). Addresses review feedback Co-authored-by: padivedi <padivedi@amd.com>
…kernels (llvm#191770) Don't use the L0 heuristics if all the dimensions are specified by the user code.
…#190026) Add translation from the MLIR OpenMP depend clause with iterator modifier to LLVM IR. `buildDependData` (in OpenMPToLLVMIRTranslation) allocates a single `kmp_depend_info` array sized to hold both locator (non-iterated) and iterated entries. Locator dependencies use the existing static path (a vector of `DependData`), while iterated dependencies use a dynamically-sized path (`DepArray`, `NumDeps`). The reason both paths are not unified under the dynamic allocation is that the existing locator path emits actual `kmp_depend_info` entries inside OMPIRBuilder methods (`createTask`, `createTarget`), whereas the iterator path must emit the iterator loop in OpenMPToLLVMIRTranslation (since the convention is to not pass MLIR ops into the OMPIRBuilder). Unifying them would require modifying existing depend clause tests. The `OMPIRBuilder::DependenciesInfo` struct is extended to hold either a `SmallVector<DependData>` (locator path) or a pre-built `{DepArray, NumDeps}` pair (iterator path). The single-entry `emitTaskDependency` helper is made public so the translation layer can fill individual `kmp_depend_info` entries inside the iterator loop body. This patch is part of the feature work for llvm#188061. Assisted with copilot.
This is a set of squashed reverts of recent clang doc patches, since its breaking something on Darwin builders: https://lab.llvm.org/buildbot/#/builders/23/builds/19172 Revert "[clang-doc][nfc] Default initialize all StringRef members (llvm#191641)" This reverts commit 155b9b3. Revert "[clang-doc] Initialize StringRef members in Info types (llvm#191637)" This reverts commit 489dab3. Revert "[clang-doc] Initialize member variable (llvm#191570)" This reverts commit 5d64a44. Revert "[clang-doc] Merge data into persistent memory (llvm#190056)" This reverts commit 21e0034. Revert "[clang-doc] Support deep copy between arenas for merging (llvm#190055)" This reverts commit c70dae8.
This PR improves native binary generation by avoiding `llvm::sys::ExecuteAndWait` call for ocloc and instead leveraging `oclocInvoke()` that consumes an in-memory SPIR-V string. Co-authored-by: Artem Kroviakov <artem.kroviakov@intel.com>
…s for any rank (llvm#188983) The fold for `vector.multi_reduction` only handled the rank-1 case with no reduction dimensions. For higher-rank vectors (e.g., `vector<2x3xf32>`) with empty reduction dims `[]`, the fold returned null, allowing `ElideUnitDimsInMultiDimReduction` to fire incorrectly. That canonicalization pattern checks that all *reduced* dims have size 1, but with zero reduction dims the check trivially passes, and the pattern then computes `acc op source` (e.g., `acc + source`) instead of the correct no-op result (`source`). This caused `--canonicalize` to produce a different value than `--lower-vector-multi-reduction` for the same program: vector.mask %m { vector.multi_reduction <add>, %src, %src [] : vector<3x3xi32> to vector<3x3xi32> } : vector<3x3xi1> -> vector<3x3xi32> * Without --lower-vector-multi-reduction: `src + src` (e.g., 2) * With --lower-vector-multi-reduction: `src` (e.g., 1) Fix the fold to return `source` for any rank when `reduction_dims` is empty. This makes the empty-dims case consistent: the operation is a noop regardless of rank, and `ElideUnitDimsInMultiDimReduction` no longer gets a chance to mishandle it. Fixes llvm#129415 Assisted-by: Claude Code
…llvm#191756) The inner CONCAT_VECTORS result type was hardcoded to MVT::v8i1, which is only correct when BitBytes == 1. Otherwise, the inner concat produces fewer elements than 8, causing an assertion failure: Assertion `(Ops[0].getValueType().getVectorElementCount() * Ops.size()) == VT.getVectorElementCount() && "Incorrect element count in vector concatenation!"' failed. Fix by computing the inner vector type dynamically based on BitBytes.
…part 42) (llvm#191751) Tests converted from test/Lower/Intrinsics: storage_size.f90, sum.f90, system_clock.f90, trailz.f90, transfer.f90
…lvm#189241) A new MoveLastSplitAxisPattern class handles the case where the last grid axis of one tensor dimension is moved to the front of another tensor dimension's split axes, e.g. [[0, 1], [2]] -> [[0], [1, 2]]. The three bugs fixed are: 1. detectMoveLastSplitAxisInResharding: compared source.back() with target.back() instead of target.front(), preventing the pattern from being detected for resharding like [[0,1],[2]] -> [[0],[1,2]]. 2. targetShardingInMoveLastAxis: axes were appended with push_back but should be inserted at the front, producing wrong split_axes order. 3. handlePartialAxesDuringResharding: a copy_if wrote results into the wrong output variable (addressed structurally by the clean implementation). Fixes llvm#136117 Assisted-by: Claude Code
…mMemory." (llvm#191851) Reverts llvm#190681 due to buildbot breakage (llvm#190681 (comment)).
Reverts llvm#191561 This is not required anymore.
…89000) When tiling a rank-0 linalg.generic op, tileUsingSCF returns an empty loops vector (rank-0 ops have no parallel dimensions and produce no scf.forall). Two call sites unconditionally accessed tilingResult.loops.front(), causing a crash: - tileToForallOpImpl: the loop normalization block was entered whenever mixedNumThreads was empty, regardless of whether any loops exist. Guard it with \!tilingResult.loops.empty(). - TileUsingForallOp::apply: tileOps.push_back was called unconditionally. Guard it with \!tilingResult.loops.empty(). Add regression tests for both the tile_sizes and num_threads paths, verifying that the linalg.generic is preserved and no scf.forall is emitted. Fixes llvm#187073 Assisted-by: Claude Code
This patch re-enables unicode tests on Windows by improving the `Terminal::SupportsUnicode` check. Checking that the stdout handle is a `FILE_TYPE_CHAR` is a better heuristic than always returning true, which assumed we were always using a terminal and never piping the output.
llvm#191835) an issue reported with this patch llvm#191241. Revert for now and reenable later This reverts commit e71da01.
This PR fixes a crash due to a failed assertion in the `from_python` implementations of the type casters. The assertion obviously only triggers if assertions are enabled, which isn't the case for many Python installations, *and* if a Python capsule of the wrong type is attempted to be used, so this this isn't triggered easily. The problem is that the conversion from Python capsules may set the Python error indicator but the callers of the type casters do not expect that. In fact, if there are several operloads of a function, the first may cause the error indicator to be set and the second runs into the assertion. The fix is to unset the error indicator after a failed capsule conversion, which is indicated with the return value of the function anyways. In alternative fix would be to unset the error indicator *inside* the `mlirPythonCapsuleTo*` functions; however, their documentations does say that the Python error indicator is set, so I assume that some callers may *want* to see the indicator and that the responsibility to handle it is on them. Signed-off-by: Ingo Müller <ingomueller@google.com>
llvm#191773) The old name was misleading because this function is not specific to unary ops suggested in llvm#189099 (comment)
…#191493) Follow-up to llvm#188113 per @erichkeane's feedback: `isFundamentalIntType` and `isFundamental()` should not disagree. The previous patch added `!isBitInt()` only inside `IntType::isFundamental()`, leaving the underlying TableGen predicates (`CIR_AnyFundamentalIntType` etc.) unaware of `_BitInt`. That meant `isSignedFundamental()` and `isUnsignedFundamental()` were silently wrong — a `_BitInt(32)` would pass them. This patch adds a `CIR_IsNotBitIntPred` to the three fundamental-int constraint defs so everything stays consistent. `isFundamental()` now just forwards to `isFundamentalIntType()` with no extra logic. Includes an `invalid-bitint.cir` test that checks a `_BitInt(32)` is rejected where a fundamental unsigned int is required. Made with [Cursor](https://cursor.com)
Add CIR-to-LLVM and classic codegen RUN lines to empty.cpp, c89-implicit-int.c, expressions.cpp, binop.c, forward-enum.c, and static-vars.c so each test verifies LLVM IR output from both pipelines. Made with [Cursor](https://cursor.com)
) PR llvm#181071 caused regressions on Linux on Arm. These are being tracked in: - llvm#191855 - llvm#191859 This PR disables the failing tests for now, to fix the broken buildbot.
MachOPlatform::HeaderOptions now includes an optional UUID field. If set, this will be used to build an LC_UUID load command for the JITDylib's MachO header. No testcase: MachOPlatform construction requires the ORC runtime, which we can't require in LLVM regression or unit tests. In the future we should test this through the ORC runtime.
…llvm#191875) Depending on the case, SLP either misses optimizing re-vectorized runtime strided loads (and use a gather instead) or produces the incorrect strided load.
…nclude in `UniqueBBID.h` (llvm#191877) The modules build of LLVM broke when this patch landed ``` commit 2f422a5 Author: Rahman Lavaee <rahmanl@google.com> Date: Fri Apr 10 15:58:16 2026 -0700 [Codegen, X86] Add prefetch insertion based on Propeller profile (llvm#166324) ``` with an error like: ``` [2026-04-11T10:33:41.699Z] While building module 'LLVM_Utils' imported from /Users/ec2-user/jenkins/workspace/m.org_clang-stage2-Rthinlto_main/llvm-project/llvm/lib/Demangle/Demangle.cpp:13: [2026-04-11T10:33:41.699Z] In file included from <module-includes>:321: [2026-04-11T10:33:41.699Z] /Users/ec2-user/jenkins/workspace/m.org_clang-stage2-Rthinlto_main/llvm-project/llvm/include/llvm/Support/UniqueBBID.h:40:3: error: missing '#include "llvm/ADT/StringRef.h"'; 'StringRef' must be declared before it is used [2026-04-11T10:33:41.699Z] 40 | StringRef TargetFunction; [2026-04-11T10:33:41.699Z] | ^ [2026-04-11T10:33:41.699Z] /Users/ec2-user/jenkins/workspace/m.org_clang-stage2-Rthinlto_main/llvm-project/llvm/include/llvm/ADT/StringRef.h:55:24: note: declaration here is not visible [2026-04-11T10:33:41.699Z] 55 | class LLVM_GSL_POINTER StringRef { [2026-04-11T10:33:41.699Z] | ^ [2026-04-11T10:33:41.699Z] /Users/ec2-user/jenkins/workspace/m.org_clang-stage2-Rthinlto_main/llvm-project/llvm/lib/Demangle/Demangle.cpp:13:10: fatal error: could not build module 'LLVM_Utils' [2026-04-11T10:33:41.699Z] 13 | #include "llvm/Demangle/Demangle.h" [2026-04-11T10:33:41.699Z] | ~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~ ``` https://ci.swift.org/job/llvm.org/job/clang-stage2-Rthinlto/job/main/150/ This patch tries to fix that by adding the missing include. rdar://174555346
…t rotations (llvm#157208) This does exactly what AArch64 does.
…rgets (llvm#187134) Currently, when `X86_64ABIInfo::classifyRegCallStructTypeImpl` classifies a struct argument or return value as direct, it leaves the LLVM IR coerce type unspecified, implicitly relying on `CodeGenTypes::ConvertType` to eventually construct a default IR type based on the struct's layout. This conversion is neither stable nor guaranteed to adhere to the ABI's classification rules. Instead, rewrite `classifyRegCallStructTypeImpl` to construct an explicit sequence of coerce types, using the existing field classification to obtain a coerce type for each member of the struct. Also, rename the function to `passRegCallStructTypeDirectly` and return a boolean instead, so that now `classifyRegCallStructType` is the only place that computes `ABIArgInfo`. This rewrite also fixes several other issues with the `X86_64ABIInfo` implementation of `__regcall`: * Empty structs are now ignored instead of being misclassified as direct. * Arrays are now classified specially based on the element type, since `X86_64ABIInfo::classifyArgumentType` ignores standalone array types. * SSE registers used for return values are now correctly reused for arguments, matching the 64-bit Windows behavior. Since this is an ABI change, it has the potential to cause incompatibilities with `__regcall` code compiled by earlier versions of Clang. Specifically: * Because SSE return registers can now be reused as argument registers, functions will now pass more floating point arguments in SSE registers. * `_Complex float` struct fields are now passed in one SSE register instead of two. Fixes llvm#62999 Fixes llvm#98635
li -1 can be compressed to c.li.
lamb-j
requested review from
Groverkss,
antiagainst,
kuhar,
nicolasvasilache and
stellaraccident
as code owners
Collaborator
|
PSDB Build Link: http://mlse-bdc-20dd129:8065/#/builders/11/builds/209 |
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 join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Automated merge of upstream llvm/llvm-project main branch.
This PR was created automatically by the upstream-merge workflow.