[flang-commits] [flang] [flang][OpenMP] Inline the firstprivate array copy instead of calling Assign() (PR #211543)

[Spencer Bryngelson via flang-commits]

sbryngelson wrote:

Closing this in favour of #209539, which landed as e949b654424b.

It fixes the same defect one layer down, in `fir::AliasAnalysis::alias`, rather than at the `InlineHLFIRAssign` call site — so the existing `aliasRes.isNo()` path inlines the copy on its own and the special case here is redundant. That is the better-layered fix: it covers every consumer of alias analysis, not one call site.

Measured on gfx90a at `d1d3891077f6`, `-O3`, one-element `real(8)` array:

| build | offload link | binary |
|---|---|---|
| baseline | `ld.lld: error: undefined symbol: _FortranAAssign` | fails |
| #209539 alone | links | 1241304 bytes |
| this PR alone | links | 1241304 bytes |

Byte-identical, so this adds nothing over what landed.

Thanks @tblah for the reviews here — the POINTER association point you raised is handled in #209539 as well, with a dedicated test.


https://github.com/llvm/llvm-project/pull/211543