[all-commits] [llvm/llvm-project] 469794: [analyzer] Recognize missing MPI collectives in MP...

[Balázs Benics via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 469794e66f86f11ae50015f6199179ca83efc58b
      https://github.com/llvm/llvm-project/commit/469794e66f86f11ae50015f6199179ca83efc58b
  Author: Balázs Benics <benicsbalazs at gmail.com>
  Date:   2026-07-15 (Wed, 15 Jul 2026)

  Changed paths:
    M clang/lib/StaticAnalyzer/Checkers/MPI-Checker/MPIFunctionClassifier.cpp
    M clang/test/Analysis/MPIMock.h
    M clang/test/Analysis/mpichecker.cpp

  Log Message:
  -----------
  [analyzer] Recognize missing MPI collectives in MPI-Checker (#209263)

Fixes a false positive in the `optin.mpi.MPI-Checker` where legitimate
code was flagged with "Request 'X' has no matching nonblocking call."

Background for reviewers unfamiliar with MPI or this checker
------------------------------------------------------------

MPI (Message Passing Interface) is the standard API for communication
between processes in HPC/cluster programs (C/C++/Fortran). Many MPI
operations come in a *nonblocking* form, prefixed with "I" (for
"immediate"): e.g. `MPI_Isend`, `MPI_Iscatter`. A nonblocking call
returns immediately and hands back an `MPI_Request` handle; the caller
must later complete it with `MPI_Wait` (or `MPI_Waitall`). Forgetting to
wait, or reusing a request before waiting, are real bugs (buffer reuse
races, leaked requests).

The MPI-Checker (originally contributed as a GSoC 2016 project, and is a
domain-specific check - hence it's in "optin").
It tracks each `MPI_Request` region through a small state machine and
reports: three misuse patterns:

* **missing wait**: a nonblocking call whose request is never waited on
* **double nonblocking**: a request reused by a second nonblocking call
before being waited on
* **unmatched wait**: `MPI_Wait` on a request with no originating
nonblocking call

Root cause
----------

`MPIFunctionClassifier` maintains a hand-written list of known MPI
functions. That list was incomplete: several standard collectives --
notably the vector variants (`MPI_Iscatterv`, `MPI_Igatherv`, ...), the
scans, the reduce-scatter family, and `MPI_Ibarrier` -- were entirely
absent. Because `isNonBlockingType()` returned `false` for them,
`checkDoubleNonblocking()` never recorded the request as Nonblocking, so
the matching `MPI_Wait` looked unmatched and produced a FP.

Note that only
`isNonBlockingType()`/`isWaitType()`/`isMPI_Wait()`/`isMPI_Waitall()`
actually drive the checker; the per-shape predicates (`isScatterType`,
`isGatherType`, `isCollToColl`, etc.) and their backing containers are
currently unused. `checkDoubleNonblocking()` also reads the request from
the *last* argument, which is generic since every nonblocking MPI
function places `MPI_Request*` last -- so no per-function argument
handling is needed.

Fix
---

Register the remaining standard collectives and mark the nonblocking
("I") variants as nonblocking:

* **vector variants**: `MPI_Scatterv`/`Iscatterv`,
`MPI_Gatherv`/`Igatherv`, `MPI_Allgatherv`/`Iallgatherv`,
`MPI_Alltoallv`/`Ialltoallv`, `MPI_Alltoallw`/`Ialltoallw`
  * **nonblocking barrier**: `MPI_Ibarrier`
  * **scans**: `MPI_Scan`/`Iscan`, `MPI_Exscan`/`Iexscan`
* **reduce-scatter**: `MPI_Reduce_scatter`/`Ireduce_scatter`,
`MPI_Reduce_scatter_block`/`Ireduce_scatter_block`

These functions are not referenced by any per-shape predicate, so they
are added directly to the classification containers without a dedicated
`IdentifierInfo` member. Neighborhood collectives (`MPI_Ineighbor_*`)
and the wait-family beyond `Wait`/`Waitall`
(`Waitany`/`Waitsome`/`Test*`) remain out of scope.

Adds regression tests covering the reported repro (no warning), and
confirms double-nonblocking and missing-wait are still detected for the
new functions.

Fixes #208107

---------

Co-authored-by: Claude Opus 4.8 <noreply at anthropic.com>



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications