Upstream compiler patches

5
patches merged: 3
bugs fixed by maintainers: 10
still open

Every change, and where it stands

Merged 5

triton #11324
Treat the warp_specialize terminators as CTA sync points
30 redundant barriers removed from an H100 build
llvm-project #216851
Fix a mem2reg crash on a zero-extent alloca
mlir-opt no longer asserts on a zero-sized memref
llvm-project #217392
Do not read non-constant loop bounds when unrolling
three shapes where the trip count answered but the bounds were not constant
llvm-project #216854
Verify multi_reduction reduction dimensions
an out-of-bounds write reachable from a malformed op
triton #11311
Make the Example 4 plugin block in the docs parse
documentation

Reported, fixed by a maintainer 3

triton #11328
Multicast TMA into a memdesc view misses its cluster barrier
fixed by Jokeren
triton #11404
A relaxed cluster_barrier is treated as a full cluster sync point
fixed by lezcano
triton #11407
predicateOp drops the predicate on NotSpeculatable arithmetic
fixed by lezcano

Open 10, five shown

triton #10766
Fold split(join(a, b)) and join(split(x))
shared memory 16,384 B to 0, SASS 224 to 72 instructions
triton #11325
Key subslice offset comparison on the value they came from
soundness fix in Membar
llvm-project #218226
Only convert an scf.yield that terminates an scf.for
a cast abort on vector.contract inside scf.if
llvm-project #216853
Do not coalesce when an inner loop reads an outer iter_arg
four more MLIR fixes are open alongside this one
sparsify #143
Shard the latent dimension across ranks
two more sparsify patches are open alongside this one

States verified against the GitHub API on 27 August 2026. The three reported issues were fixed by Triton maintainers, not by a patch of mine: the credit there is for the report and the reproducer.

The Membar patch removes about 30 `bar.sync` instructions and buys no speedup

Modelling ttg.warp_yield and ttg.warp_return as a CTA-wide rendezvous stops Triton's Membar analysis inserting a redundant barrier after every warp_specialize. On an H100 build the generated PTX loses about 30 bar.sync instructions and runs somewhere between 0.5 percent slower and 0.3 percent faster. I published that negative result on the pull request instead of quoting the barrier count and leaving it to imply a win.

Three MLIR patches: a crash, an unchecked read, a buffer overrun

Patch	What was wrong
`mem2reg` on a zero-extent alloca	Promotion never checked the extent in the memref shape, so `mlir-opt --mem2reg` asserted on legal input.
`scf` loop unroller	`loopUnrollByFactor` read bounds it had not established were constant, on the shapes where the trip-count query answered anyway.
`vector.multi_reduction` verifier	Out-of-range and duplicate `reduction_dims` reached `getReductionMask()`, which wrote past the end of a `SmallVector<bool>`.

A documentation fix sat 43 days, then merged in under a day

I first sent it inside a larger pull request, next to a change a maintainer had already pushed back on, and it waited 43 days without CI. Extracted onto a fresh branch, one file, nothing contested, it merged in under a day. I no longer tie an uncontroversial fix to a contested one.

Four changes did not land, and I withdrew one myself

Change	Outcome
Membar twin on the entry side	Declined as marginal. I had measured that it left the generated code unchanged.
fp8 block-pointer loads	Declined: block pointers are deprecated and being removed.
Defensive error path in `global_scratch_alloc`	Withdrawn after a reviewer showed the assert is unreachable from generated code.
Per-lane retirement pass (RFC)	Maintainers asked for it out of tree. It ships as a loadable plugin.

Every state on this page was checked against the GitHub API on 27 August 2026

That covers the merged patches, the ten changes still open under review, and the closures above. Where maintainers fixed a reported bug themselves from my reproducer, the page credits the report.