llvm-project
255d0013 - [flang][OpenMP] Don't emit implicit default mapper for flat allocatable derived types (#209645)

Commit
119 days ago
[flang][OpenMP] Don't emit implicit default mapper for flat allocatable derived types (#209645) An allocatable or pointer capture of a derived type in a `target` region triggers synthesis of an implicit default declare mapper for the type. The gate keyed only on whether the captured *variable* was allocatable, not on whether the *type* needs a mapper: ```cpp if (!isPointer && (hasDefaultMapper || isAllocatable)) { ``` For a flat type -- all components trivially mappable, no allocatable/pointer/nested record members -- the mapper is unnecessary. The object maps as a plain bulk copy. Emitting a mapper anyway lowers to a per-element, per-component mapper walk in the offload runtime (`targetDataBegin` -> `targetDataMapper` -> `targetDataBegin`, and the mirror walk in `targetDataEnd`), which is `O(num_elements * num_components)` per kernel invocation. On a large device-resident allocatable array this busy-loops the host for minutes with the GPU idle. `Fortran::lower::omp::requiresImplicitDefaultDeclareMapper()` already answers whether a type requires a mapper (allocatable/pointer/nested-record components, or ISO-C interop), but the allocatable-capture path never consulted it. This gates on it: ```cpp if (!isPointer && (hasDefaultMapper || (isAllocatable && requiresImplicitDefaultDeclareMapper(*typeSpec)))) { ``` Pointer captures and user/pre-existing declare mappers are unaffected. A flat allocatable now maps as descriptor + base-address + attach entries with no `mapper()` attribute -- identical to the pointer path. ### Reproducing the runtime cost Reported against the AMD ROCm flang runtime as ROCm/llvm-project#3385. A flat derived type (`ghost_point`-like) in a device-resident allocatable array, mapped implicitly across many kernel invocations, hangs for minutes. `perf` on the hang shows all time in per-component work: `DenseMap` insert into the mapping-state tables, `targetDataEnd`'s mirror walk, and `SourceInfo` construction -- each `O(1)`/`O(log M)` per entry, so the total is linear in `M = num_elements * num_components * kernel_invocations` with a large constant. Suppressing the mapper collapses `M` to one bulk entry. ### Tests - `implicit-map-flat-allocatable-no-mapper.f90`: new; a flat allocatable capture emits no mapper. - `implicit-map-pointer-no-default-mapper.f90`: the type now has an allocatable component so it still requires a mapper (keeps the pointer-vs-allocatable distinction meaningful). - `derived-type-map.f90`, `defaultmap.f90`: updated; these asserted a mapper for a flat allocatable capture, which is the behavior being removed. ### AI tool usage Per the [LLVM AI Tool Use Policy](https://llvm.org/docs/AIToolPolicy.html): this change was developed with substantial assistance from Claude (Opus 4.8) and reviewed by the author. The root-cause analysis, the patch, the tests, and this description were AI-assisted; the reproducer and the offload-runtime `gdb`/`perf` profiling on an AMD Instinct MI250X (HPC Fund) that confirmed the runtime cost is linear (not a runtime bug) were part of that workflow. Also noted as an `Assisted-by:` trailer in the commit.
Author
Parents
Loading