[AIROCMLIR-1056] LLVM Upstream Merge - June...August 2026

[mirza-halilcevic]

Motivation

LLVM upstream bump for June-August 2026.

Technical Details

Build fixes

• Bufferization API (external/llvm-project/.../Bufferize.cpp): Update the unknownTypeConverterFn lambda in getPartialBufferizationOptions() to the new upstream signature (TensorLikeType → BufferLikeType instead of TensorType → MemRefType), matching the other updated call sites.
• MFMAPermB namespace move (MfmaInsnGroup.{h,cpp}): Upstream moved the MFMAPermB enum from the amdgpu namespace to ROCDL. Updated the field type, initializers, and using declarations, and added the ROCDLDialect.h include.
• DenseMapInfo refactor (RockPipeline.cpp, BufferLoadMerge.cpp): Upstream reworked DenseMap to track occupied buckets via a separate bit-vector, so DenseMapInfo no longer exposes getEmptyKey()/getTombstoneKey() and now auto-provides an enum specialization. Removed the now-redundant custom DenseMapInfo<MemoryAccessType> specialization and dropped the empty/tombstone-key guard from SimpleOperationInfo::isEqual (mirroring the upstream CSE change).
• splitFlatOffset signature (external/llvm-project/.../SIRegisterInfo.cpp): Upstream replaced the uint64_t FLAT discriminant with the scoped AMDGPU::FlatAddrSpace enum (llvm/llvm-project#200431), so the SIInstrFlags::FlatScratch argument in the SVS constant-offset fold no longer converts. Pass AMDGPU::FlatAddrSpace::FlatScratch instead. Note this arrives late because [EXTERNAL][AMDGPU] Fold constant offset into SV form addressing in SVS #2435 landed on develop on 2026-08-03, after this branch's June subtree pull.
• APFloat::convert inexact status for zero-less / sign-less formats (Rock/utility/builderUtils.{h,cpp}, rocmlir-gen.cpp): Upstream's APFloat::convert now reports opInexact (and snaps to the nearest representable value) when the target format has no encoding for the value; i.e. zero for Float8E8M0FNU (!hasZero) or negatives for sign-less formats (!hasSignedRepr). createConstantFloatOp/createFloatSplatTensor asserted status == opOK, so building a zero constant of an E8M0 scale operand (e.g. the scaled-GEMM input fill in rocmlir-gen) aborted with an assertion failure, cascading into most check-rocmlir failures. Added isAcceptableConversionStatus() to tolerate that unavoidable inexactness while still catching genuine precision loss.
• Added bufferization::registerBufferizableOpInterfaceExternalModels(registry) in InitRocMLIRDialects.h.
• Dropped MLIRNVGPUDialect from librockcompiler_deps.cmake.

Test fixes

• Lit internal-shell migration (mlir/test/lit.cfg.py, mlir/test/e2e/lit.cfg.py, mlir/test/fusion/e2e/lit.cfg.py): ShTest(execute_external=True) is removed in upstream lit. Switched all three configs to default to the internal shell (honoring LIT_USE_INTERNAL_SHELL), matching upstream MLIR's own lit.cfg.py.
• Pipeline dump expectations (rocmlir-driver/pipelines.mlir, runner-pipelines.mlir): Refresh -dump-pipelines FileCheck expectations for new upstream pass options/passes: linalg-morph-ops's new category-to-named option, and convert-vector-to-llvm's new enable-gep-inbounds-nuw option.
• ROCDL buffer op cache-policy operand (prepare_llvm.mlir, add_alias_info.mlir): Upstream changed rocdl.raw.ptr.buffer.store and rocdl.raw.ptr.buffer.load.lds so the trailing cache-policy operand is now a cache-policy-family keyword or a 32-bit integer literal instead of an SSA value. Updated the test IR accordingly.
• MemRef strict-properties alignment (mlir/test/**, incl. CAPI/reduce_fusible.cpp): The MemRef dialect now uses strict properties in its assembly format, so the inherent alignment attribute can no longer be parsed from the discardable attr-dict. Rewrote memref.alloc/memref.alloca/memref.global from {alignment = N : i64} to the custom-format alignment = N (keeping any other discardable attrs), in both IR and CHECK lines.
• amdgpu.raw_buffer_* mandatory boundsCheck keyword (emulate-narrow-type.mlir, buffer-load-merge.mlir, lowering_global_load_store.mlir, buffer_atomic_emulation.mlir, ...): boundsCheck(<bool>) is now a required leading keyword for raw_buffer_load/store/atomic_* instead of an attr-dict entry. Added boundsCheck(true) for the default cases and converted {boundsCheck = false} to boundsCheck(false).
• BitEnumAttr quoting for features (rock-gemm-mixed-input-types.mlir, gridwise_attention_accel_lowering{,_gqa}.mlir, fold_broadcast.mlir, integration/multibuffer/test_multi_buffer_full.mlir, many fusion/*): Upstream now prints multi-bit and none bit-enum values quoted (features = "mfma|dot|atomic_add", features = "none"), and the bare pipe-separated form no longer parses. Quoted features = ... on rock.gemm/rock.gridwise_attention_accel/rock.blockwise_gemm_accel in IR and CHECK lines (single-bit values like dot stay bare, matching the printer).
• rocdl.load.to.lds / rocdl.raw.ptr.buffer.load.lds alias attrs as properties (add_alias_info.mlir): these ops now emit alias_scopes/noalias_scopes in the <...> properties syntax instead of the {...} attr-dict. Updated the CHECK lines (separate from the cache-policy-operand change above).
• rocdl.make.buffer.rsrc numRecords type (prepare_llvm.mlir): the op now prints/parses the numRecords operand type in its signature, so : !llvm.ptr to <8> became : !llvm.ptr, i64 to <8>.
• amdgpu.wmma clamp flag and empty attr-dict spacing (lowering_wmma_gemm.mlir, lowering_xdlops_gemm.mlir): clamp is now an oilist keyword (... + %c clamp : instead of ... {clamp}), and amdgpu.mfma/wmma with no discardable attrs print no attr-dict, so CHECK patterns like + [[c]] {{.*}} : were tightened to + [[c]]{{.*}} :.
• Attribute print-order changes (num_cu.mlir, waves_per_eu.mlir, quant-dot-scaled-to-matmul-t-block-scaled.mlir): inherent attributes that became properties now print after the discardable attr-dict; gpu.func's known_block_size/known_grid_size moved to the end, and tosa.matmul_t_block_scaled's block_size now prints after perf_config. Reordered the affected CHECK-SAME lines (no value changes).

Diff Files

• diff -rup llvm-project/llvm rocMLIR/external/llvm-project/llvm
  llvm-diff.txt
• diff -rup llvm-project/mlir rocMLIR/external/llvm-project/mlir
  mlir-diff.txt

Test Plan

External Tests

• check-llvm
• check-mlir

CI

• PR
• Nightly
• Weekly

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of files (300). Try reducing the number of changed files and requesting a review from Copilot again.

[rocmlir-pr-reviewer]

Verdict: APPROVE -- submitted as COMMENT (automated reviews are advisory)  ·  Findings: 0 (0 Critical, 0 Major, 0 Minor)

---

Scope

LLVM upstream merge (June 2026). The vast majority of the diff is vendored external/llvm-project changes (out of scope for rocMLIR review). The rocMLIR-side changes are 12 files: build fixes to MfmaInsnGroup.{h,cpp}, BufferLoadMerge.cpp, RockPipeline.cpp; test-infra updates to three lit.cfg.py files; FileCheck expectation refreshes in four .mlir tests; and a regenerated librockcompiler_deps.cmake.

Findings

No blocking issues found in the rocMLIR-side changes.

• MfmaInsnGroup.h/.cpp: clean amdgpu::MFMAPermB → ROCDL::MFMAPermB migration; the new mlir/Dialect/LLVMIR/ROCDLDialect.h include is placed in correct lexicographic order after AMDGPUDialect.h, and every using ROCDL::MFMAPermB; / MFMAPermB::none site is consistent.
• BufferLoadMerge.cpp / RockPipeline.cpp: removal of the custom DenseMapInfo<MemoryAccessType> specialization and of the empty/tombstone-key guard in SimpleOperationInfo::isEqual correctly tracks the upstream DenseMap rework. MemoryAccessType values (READ=1, WRITE=2, UNKNOWN=3) don't collide with the auto-provided enum specialization's sentinels, so the DenseSet<pair<GpuAllocOp, DependencyType>> usage remains sound.
• The three lit.cfg.py files default to the internal shell while honoring LIT_USE_INTERNAL_SHELL (ShTest(execute_external=not use_lit_shell)), matching upstream MLIR's own config after execute_external=True removal.
• librockcompiler_deps.cmake is the regenerated dependency list, satisfying the checklist item requiring it be updated when dependencies change.

Notes

• The .mlir FileCheck refreshes (pipelines.mlir, runner-pipelines.mlir, prepare_llvm.mlir, add_alias_info.mlir) are mechanical expectation updates for new upstream pass options and the ROCDL buffer-op cache-policy operand form; correctness rests on check-mlir passing.
• Pre-existing (out of scope): MfmaInsnGroup.cpp has no SPDX license header, and RockPipeline.cpp uses std::set/std::vector/std::map where LLVM ADTs are usually preferred — neither introduced by this PR.
• Recent commit history shows external/llvm-project changes carried under a separate [EXTERNAL]-prefixed commit, consistent with the repository convention.

CI status

No failing or cancelled checks in the pre-fetched status (only detect passing and the auto-review pipeline in progress). The unchecked PR/Nightly/Weekly boxes in the description are the author's manual checklist, not CI signal.

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of files (300). Try reducing the number of changed files and requesting a review from Copilot again.

[rocmlir-pr-reviewer]

Verdict: COMMENT  ·  Findings: 2 (0 Critical, 0 Major, 2 Minor)

---

Scope

June 2026 LLVM upstream bump. The vast majority of the diff is the external/llvm-project subtree pull, which I treated as vendored upstream content and did not review line-by-line. I reviewed the 12 rocMLIR-proper files the merge required:

• C++ build fixes — MfmaInsnGroup.{h,cpp} (amdgpu::MFMAPermB → ROCDL::MFMAPermB), RockPipeline.cpp (drop custom DenseMapInfo<MemoryAccessType>), BufferLoadMerge.cpp (drop empty/tombstone guard in SimpleOperationInfo::isEqual).
• Test fixes — three lit.cfg.py internal-shell migrations, -dump-pipelines expectation refreshes, ROCDL buffer-op cache-policy operand updates.
• Build deps — librockcompiler_deps.cmake drops MLIRNVGPUDialect.

Findings

Two Minor cleanups; nothing blocking. See inline comments on mlir/test/Dialect/Rock/add_alias_info.mlir:38 and mlir/include/mlir/Dialect/Rock/IR/MfmaInsnGroup.h:18.

Notes

Spot-checks that came back clean:

• I verified the DenseMap rework against the vendored header: external/llvm-project/llvm/include/llvm/ADT/DenseMapInfo.h no longer declares getEmptyKey()/getTombstoneKey(), and lines 138-149 now provide the auto enum specialization. So both the RockPipeline.cpp and BufferLoadMerge.cpp deletions are required, and the new enum info hashes MemoryAccessType through uint32_t exactly as the removed custom one did — no behavior change.
• All three lit.cfg.py files already have import os and import lit.util at the top, so the new lit.util.pythonize_bool call resolves; the block matches upstream MLIR's own config verbatim.
• The refreshed pass-option strings (category-to-named, enable-gep-inbounds-nuw) are inserted in the correct alphabetical slot that -dump-pipelines emits.

Out of scope, for a possible follow-up (pre-existing lines, not touched here):

• MfmaInsnGroupSelectKeyInfo in MfmaInsnGroup.h:96-104 still defines getEmptyKey()/getTombstoneKey(), which are now dead under the new DenseMap — the same dead code this PR removed from RockPipeline.cpp.
• SimpleOperationInfo in BufferLoadMerge.cpp:40 now overrides both members of its llvm::DenseMapInfo<Operation *> base, so the inheritance is vestigial.

One behavioral note worth confirming rather than a defect: on Linux the lit default flips from the external (bash) shell to lit's internal shell for all three suites. That is upstream-mandated and matches upstream MLIR, but it is a real change for E2E RUN lines, so the premerge lit run is the thing to watch.

CI status

No check is in the fail or cancel bucket. C/C++ premerge checks is still IN_PROGRESS — for a bump of this size that job is the primary correctness signal (in particular for the ROCDL::MFMAPermB type change flowing into MFMA emission, which I could not verify statically), so please confirm it goes green before merging.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #2427      +/-   ##
===========================================
+ Coverage    82.57%   83.71%   +1.14%     
===========================================
  Files          120      121       +1     
  Lines        42852    43386     +534     
  Branches      7110     8288    +1178     
===========================================
+ Hits         35381    36318     +937     
+ Misses        4815     4534     -281     
+ Partials      2656     2534     -122     

Flag	Coverage Δ
gfx120x	83.55% <ø> (+1.02%)	⬆️
gfx950	83.55% <ø> (+1.20%)	⬆️
mfma	83.36% <ø> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.
see 52 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum diff size. Try reducing the number of changed files and lines, and requesting a review from Copilot again.

[rocmlir-pr-reviewer]

Verdict: COMMENT  ·  Findings: 3 (0 Critical, 0 Major, 3 Minor)

---

Scope

LLVM upstream subtree bump into external/llvm-project/ (44,425 files changed; 44,302 of them under external/). I line-reviewed the 123 rocMLIR-proper files under mlir/ — build fixes (Bufferization external-model registration, MFMAPermB namespace move, DenseMapInfo refactor, APFloat::convert inexact-status handling) and mechanical test updates (lit internal-shell migration, MemRef alignment custom format, boundsCheck(...) keyword, BitEnumAttr quoting, ROCDL alias attrs as properties, AMDGPU MIR triple spelling). The upstream subtree bump itself was not line-reviewed.

Findings

Three Minor items, none blocking — details inline:

• mlir/lib/Dialect/Rock/utility/builderUtils.cpp:81 — exact != comparison against a bitmask-valued APFloat::opStatus.
• mlir/test/Dialect/Rock/add_alias_info.mlir:38 — constant %4 left dead after the cache-policy operand change.
• mlir/lib/Dialect/Rock/Pipelines/Pipelines.cpp:292 — assignments are now no-ops given the upstream default flip, which also silently changes the host runner pipeline.

Notes

• The PR description says eliminate-frame-index-flat-scratch-svs.mir renames prolog-epilog → prologepilog. The diff does not do that, and prolog-epilog is in fact the correct registered pass argument (MachinePassRegistry.def) — the actual hunk only rewrites the target triple to the new canonical amdgpu9.50-amd-amdhsa spelling, matching sibling AMDGPU tests. The code is right; only the description is inverted. Worth fixing the description so future archaeology isn't misled.
• Spot-checked the mechanical test rewrites and they look semantics-preserving: boundsCheck(true) matches the previous default, buffer-load-merge.mlir still contrasts boundsCheck(false)/(true) so the negative case survives, and the mfma/wmma CHECK edit [[c]] {{.*}} : → [[c]]{{.*}} : tightens rather than loosens the match.
• Two changes aren't mentioned in the description: the bufferization::registerBufferizableOpInterfaceExternalModels call added in mlir/include/mlir/InitRocMLIRDialects.h, and dropping MLIRNVGPUDialect from librockcompiler_deps.cmake. Both look intentional, but they'd be easy to lose track of on a bisect.
• @pabloantoniom's question on mlir/test/rocmlir-driver/runner-pipelines.mlir:7 (why mhal-prefill, mhal-emulate-narrow-type, and lower-affine newly appear in the dumped pipeline) is still open; I have no additional information beyond what's in the diff.

CI status

No product CI failures. Build and Test, C/C++ premerge checks, Windows component build, Python format/lint, and Jenkins all pass. The only fail entries in checks.json are this auto-review pipeline's own review job, which is expected and is not a CI failure. MIGraphX, Code coverage, and the perf/tuning jobs are still pending — given the blast radius of a subtree bump, those are worth confirming green before merge.

[rocmlir-pr-reviewer]

Verdict: REQUEST_CHANGES -- submitted as COMMENT (automated reviews are advisory)  ·  Findings: 1 (0 Critical, 1 Major, 0 Minor)

---

Scope

LLVM upstream merge covering June–August 2026. The overwhelming majority of the diff is the external/llvm-project subtree bump; the rocMLIR-side changes are a small set of source fixes (API renames, removed DenseMapInfo specializations, a new APFloat conversion-status helper, dialect registration, a CMake dependency drop, three lit.cfg.py internal-shell migrations) plus a large mechanical refresh of .mlir FileCheck expectations for upstream syntax changes (memref alignment = custom format, quoted BitEnumAttr values, boundsCheck(...) keyword, ROCDL alias attrs as properties, attribute print ordering).

Findings

One Major issue: mlir/test/rocmlir-driver/runner-pipelines.mlir:17 expects include-min-max-f=false include-min-max-i=false, which directly contradicts external/mlir-hal/lib/Dialect/MHAL/Pipelines/Pipelines.cpp:109-110 where this same PR sets both options to true. The test will fail whenever it is actually executed.

Notes

• The external/mlir-hal change is correctly isolated in its own [EXTERNAL]-prefixed commit.
• rock::isAcceptableConversionStatus (mlir/lib/Dialect/Rock/utility/builderUtils.cpp) was reviewed closely: it masks against expectedStatus and tolerates only an extra opInexact bit, gated on the target semantics genuinely lacking a zero or signed encoding. A real precision loss (e.g. 3.0f into Float8E8M0FNU) still asserts. Looks correct.
• Removal of the hand-written DenseMapInfo<MemoryAccessType> in RockPipeline.cpp is safe — MemoryAccessType is a scoped enum with an underlying uint32_t, which upstream now auto-specializes. Likewise the getEmptyKey()/getTombstoneKey() checks dropped from BufferLoadMerge.cpp's isEqual are no longer reachable under the new occupied-bucket representation.
• The three lit.cfg.py rewrites all have import os and import lit.util already in scope, and mirror the upstream pattern.

Answering the open question on runner-pipelines.mlir:7

mhal-prefill, mhal-emulate-narrow-type, and the second lower-affine are not newly added by this PR — all three are already present in the host runner pipeline construction on develop (external/mlir-hal/lib/Dialect/MHAL/Pipelines/Pipelines.cpp:81-91). This test carries // REQUIRES: rocm-runner on line 1, so it is skipped in default CI, and its expectations had simply drifted out of date. This PR is bringing them back in sync. The same gate is why the defect flagged above would otherwise go unnoticed.

CI status

Red at review time: Jenkins (FAILURE), the internal ML CI status check (ERROR), and Windows Component Build (ERROR). Green: C/C++ premerge checks, Build and Test, Python format and lint checks, Python performance script tests. Worth confirming the Windows and Jenkins failures are unrelated to the pipeline-option change above.