[Mlir-commits] [mlir] [mlir][XeGPU][Transform] Add gather/scatter coalescing analysis. (PR #201684)

Md Abdullah Shahneous Bari llvmlistbot at llvm.org
Wed Jun 24 08:57:45 PDT 2026


https://github.com/mshahneo updated https://github.com/llvm/llvm-project/pull/201684

>From 1584887ae15b2dd831bf64630ed77a4a7f90dd68 Mon Sep 17 00:00:00 2001
From: "Shahneous Bari, Md Abdullah" <md.abdullah.shahneous.bari at intel.com>
Date: Thu, 4 Jun 2026 18:04:30 +0000
Subject: [PATCH 1/7] [mlir][XeGPU][Transform] Add gather/scatter coalescing
 analysis.

Adds an AxisInfo-based dataflow analysis that classifies xegpu.load /
xegpu.store gather/scatter ops as coalescible and stamps a
`#xegpu.coalesce_hint<factor = N>` discardable attribute on them. The
analysis performs no rewrite of its own; the hint is consumed by a later
layout-propagation step (subsequent PRs) or by the apply helpers
(applyCoalesceGatherScatterHint) for testing.

Accesses tied to a vector.multi_reduction are not coalesced (isReductionTied):
consuming a coalesced lane_data[FCD] on a reduction needs reduction-aware
layout handling added in follow-up PRs; without it a coalesced store fed by a
reduction would force an unlowerable lane_data convert. Pure elementwise
gather/scatter is unaffected.

Includes:
  - CoalesceHintAttr (XeGPUAttrs.td, XeGPUDialect.cpp).
  - runCoalesceGatherScatterAnalysis / applyCoalesceGatherScatterHint(s) /
    clearCoalesceGatherScatterHints public APIs (Transforms.h,
    XeGPUCoalesceGatherScatter.cpp).
  - A test-only pass (--test-xegpu-coalesce-gather-scatter, with
    analyze-only) exercising the analysis and apply.
  - Lit tests coalesce-gather-scatter{,-analyze}.mlir.

Co-Authored-By: Claude Opus 4.8 <noreply at anthropic.com>
---
 .../mlir/Dialect/XeGPU/IR/XeGPUAttrs.td       |   41 +
 .../Dialect/XeGPU/Transforms/Transforms.h     |   44 +
 mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp    |   15 +
 .../Dialect/XeGPU/Transforms/CMakeLists.txt   |    1 +
 .../Transforms/XeGPUCoalesceGatherScatter.cpp | 1171 +++++++++++++++++
 .../coalesce-gather-scatter-analyze.mlir      |  105 ++
 .../XeGPU/coalesce-gather-scatter.mlir        |  476 +++++++
 .../lib/Dialect/XeGPU/TestXeGPUTransforms.cpp |   43 +
 8 files changed, 1896 insertions(+)
 create mode 100644 mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
 create mode 100644 mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir
 create mode 100644 mlir/test/Dialect/XeGPU/coalesce-gather-scatter.mlir

diff --git a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
index 40edce8a60429..50b67cefafa31 100644
--- a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
+++ b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
@@ -927,6 +927,47 @@ def XeGPU_MemLayoutAttr : XeGPUAttr<"MemLayout", "mem_layout"> {
 
 }
 
+def XeGPU_CoalesceHintAttr : XeGPUAttr<"CoalesceHint", "coalesce_hint"> {
+  let summary = [{Per-op hint stamped by the coalesce-gather-scatter analysis.}];
+
+  let description = [{
+    `CoalesceHintAttr` is a discardable attribute attached by the
+    coalesce-gather-scatter analysis to `xegpu.load` / `xegpu.store` ops
+    whose offsets describe a contiguous-per-lane access pattern. It records
+    the chosen `lane_data` factor along the fastest-changing dim (FCD) of
+    the value vector. The `lane_layout` along the FCD is *not* stored here;
+    it is re-derived at apply time from the op's offsets inner extent and
+    the chip's subgroup size, mirroring `xegpu-propagate-layout`'s default
+    rule.
+
+    The attribute is consumed and removed by `applyCoalesceGatherScatterHint`,
+    which installs an equivalent `xegpu.layout` on the op. A producer that
+    decides not to apply the hint (for example, a propagator that detects a
+    conflict with an anchor-driven layout) should remove the attribute
+    rather than leave it dangling.
+
+    Example:
+    ```mlir
+    %v = xegpu.load %ptr[%offsets], %mask
+            <{chunk_size = 1 : i64,
+              xegpu.coalesce_hint = #xegpu.coalesce_hint<factor = 4>}>
+            : i64, vector<64xindex>, vector<64xi1> -> vector<64xf32>
+    ```
+  }];
+
+  let parameters = (ins "IntegerAttr": $factor);
+
+  let builders = [
+    AttrBuilder<(ins "int64_t":$factor), [{
+      return $_get($_ctxt, IntegerAttr::get(IntegerType::get($_ctxt, 64),
+                                            factor));
+    }]>
+  ];
+
+  let assemblyFormat = "`<` `factor` `=` $factor `>`";
+  let genVerifyDecl = 1;
+}
+
 def AnchorLayoutInterface : OpInterface<"AnchorLayoutInterface"> {
   let cppNamespace = "::mlir::xegpu";
 
diff --git a/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h b/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h
index 919a69908bdce..6046f78dd59c0 100644
--- a/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h
+++ b/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h
@@ -80,6 +80,50 @@ void populateXeGPUSgToLaneDistributeTypeConversionAndLegality(
     TypeConverter &typeConverter, RewritePatternSet &patterns,
     ConversionTarget &target);
 
+//===----------------------------------------------------------------------===//
+// Coalesce gather/scatter analysis + apply.
+//===----------------------------------------------------------------------===//
+
+/// Discardable attribute name carrying the coalesce hint
+/// (`#xegpu.coalesce_hint<factor = N>`) stamped by
+/// `runCoalesceGatherScatterAnalysis`.
+inline StringRef getCoalesceHintAttrName() { return "xegpu.coalesce_hint"; }
+
+/// Options controlling `runCoalesceGatherScatterAnalysis`.
+struct CoalesceGatherScatterAnalysisOptions {
+  /// Upper bound on the per-lane chunk size produced by coalescing. Mirrors
+  /// the `max-chunk-size` option of the original pass.
+  unsigned maxChunkSize = 8;
+};
+
+/// Run the AxisInfo-based coalescing analysis over `root` and stamp a
+/// `xegpu.coalesce_hint` attribute on every `xegpu.load` / `xegpu.store`
+/// the analysis classifies as coalescible. Ops with an existing non-empty
+/// `lane_data`, an explicit `chunk_size > 1`, or a non-uniform mask are
+/// skipped (no hint stamped).
+///
+/// This function performs no rewrite of its own; the hint is consumed by
+/// `applyCoalesceGatherScatterHint` (or a downstream pass that wants
+/// stronger control over when to honor the hint).
+void runCoalesceGatherScatterAnalysis(
+    Operation *root, const CoalesceGatherScatterAnalysisOptions &options = {});
+
+/// Apply a stamped `xegpu.coalesce_hint` on `op`: install an equivalent
+/// `lane_layout` / `lane_data` / `inst_data` layout, drop a trivial
+/// `chunk_size = 1` attribute, and remove the hint. Idempotent — no-op if
+/// the op carries no hint. Returns `failure()` when the hint is malformed
+/// (e.g. attached to an op that isn't a gather/scatter).
+LogicalResult applyCoalesceGatherScatterHint(Operation *op);
+
+/// Walk `root` and apply coalesce hints on every op that carries one.
+/// Hints stamped on unsupported ops are silently dropped.
+void applyCoalesceGatherScatterHints(Operation *root);
+
+/// Walk `root` and remove any leftover `xegpu.coalesce_hint` attributes —
+/// useful as a cleanup after a propagator has decided whether to honor each
+/// hint.
+void clearCoalesceGatherScatterHints(Operation *root);
+
 /// Collect a set of patterns to unroll xegpu operations to a smaller shapes.
 /// Users can control whether an operation to be unrolled or not, as well as
 /// its target shape via `options` structure. (via setting filterConstraint
diff --git a/mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp b/mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp
index e92b109c2223e..8cdef2dd994d8 100644
--- a/mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp
+++ b/mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp
@@ -1213,6 +1213,21 @@ RangeAttr::verify(llvm::function_ref<mlir::InFlightDiagnostic()> emitError,
   return success();
 }
 
+//===----------------------------------------------------------------------===//
+// XeGPU_CoalesceHintAttr
+//===----------------------------------------------------------------------===//
+
+LogicalResult CoalesceHintAttr::verify(
+    llvm::function_ref<mlir::InFlightDiagnostic()> emitError,
+    IntegerAttr factor) {
+  int64_t f = factor.getInt();
+  if (f < 2)
+    return emitError() << "'factor' : " << f << " must be >= 2";
+  if ((f & (f - 1)) != 0)
+    return emitError() << "'factor' : " << f << " must be a power of two";
+  return success();
+}
+
 //===----------------------------------------------------------------------===//
 // XeGPU_TensorDescType
 //===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/CMakeLists.txt b/mlir/lib/Dialect/XeGPU/Transforms/CMakeLists.txt
index 37922f7ef7d24..1b013400cff34 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/CMakeLists.txt
+++ b/mlir/lib/Dialect/XeGPU/Transforms/CMakeLists.txt
@@ -1,6 +1,7 @@
 add_mlir_dialect_library(MLIRXeGPUTransforms
   XeGPUArrayLengthOptimization.cpp
   XeGPUBlocking.cpp
+  XeGPUCoalesceGatherScatter.cpp
   XeGPUSgToLaneDistribute.cpp
   XeGPUUnroll.cpp
   XeGPUWgToSgDistribute.cpp
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
new file mode 100644
index 0000000000000..669794f5c41fd
--- /dev/null
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
@@ -0,0 +1,1171 @@
+//===- XeGPUCoalesceGatherScatter.cpp - Coalesce scatter accesses --------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass coalesces neighbouring lanes of `xegpu.load` / `xegpu.store` ops
+// so that each lane handles `N` contiguous elements along the innermost
+// dimension. The decision is driven by a small XeGPU-local axis-info
+// dataflow analysis modeled on Triton's `AxisInfo` (`contiguity`,
+// `constancy`, `divisibility`) and is applied by attaching a
+// `lane_data` layout to the original op. The actual memory-message rewrite
+// is left to the downstream WG-to-SG / SG-to-Lane distribution passes,
+// which interpret `lane_data`.
+//
+// The analysis tracks per-axis information for vectors of integer / index
+// type at any rank. The coalescing decision is computed against the
+// innermost dimension. 2-D offsets vectors with a leading unit dimension
+// (e.g. `vector<1x32xindex>`) are handled by treating the inner dim as the
+// lane dim.
+//
+// All-equal offsets ("uniform inner dim") are detected by the analysis
+// but the pass currently leaves such ops alone — there is no layout-only
+// encoding for "all lanes load the same scalar", and the previous
+// length-1-load + `vector.broadcast` rewrite was removed because it
+// conflicts with downstream layout propagation.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Analysis/DataFlow/DeadCodeAnalysis.h"
+#include "mlir/Analysis/DataFlow/SparseAnalysis.h"
+#include "mlir/Analysis/DataFlowFramework.h"
+#include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/Dialect/Vector/IR/VectorOps.h"
+#include "mlir/Dialect/XeGPU/IR/XeGPU.h"
+#include "mlir/Dialect/XeGPU/Transforms/Passes.h"
+#include "mlir/Dialect/XeGPU/Transforms/Transforms.h"
+#include "mlir/Dialect/XeGPU/Utils/XeGPUUtils.h"
+#include "mlir/Dialect/XeGPU/uArch/IntelGpuXe2.h"
+#include "mlir/IR/BuiltinAttributes.h"
+#include "mlir/IR/BuiltinTypes.h"
+#include "mlir/IR/Matchers.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+#include "llvm/ADT/APInt.h"
+#include "llvm/ADT/bit.h"
+#include "llvm/Support/MathExtras.h"
+#include <numeric>
+#include <optional>
+
+#define DEBUG_TYPE "xegpu-coalesce-gather-scatter"
+
+using namespace mlir;
+
+// AxisInfo and AxisInfoAnalysis are intentionally placed in a named namespace
+// (not anonymous) so the `dataflow::Lattice<AxisInfo>` template instantiation
+// gets a stable, externally-visible name. The TypeID machinery requires that
+// for `MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID` to apply.
+namespace mlir::xegpu::detail::axis_dataflow {
+
+//===----------------------------------------------------------------------===//
+// AxisInfo: per-axis contiguity / constancy / divisibility lattice.
+//===----------------------------------------------------------------------===//
+
+/// Sentinel "very large" value for unconstrained dimensions. Any real shape
+/// is far smaller, so component-wise `min` will collapse this to the truth.
+static constexpr int64_t kAxisInfoTop = 1LL << 30;
+
+/// Per-dimension axis information for an SSA value of integer / index type.
+///   - `contiguity[d]`: the largest N such that consecutive lanes along
+///     dimension `d` differ by exactly 1 in runs of length `N` (lane-stride
+///     1 contiguity).
+///   - `constancy[d]`: the largest N such that consecutive lanes along
+///     dimension `d` are all equal in runs of length `N`.
+///   - `divisibility[d]`: a power-of-two divisor of every element along
+///     dimension `d`.
+///   - `knownConstant`: scalar value if the entire vector is uniformly
+///     known to be a single constant.
+///   - `innerStride`: when set, every "row" along the innermost dimension
+///     is an arithmetic progression with this stride. Per-row base may
+///     differ across outer indices; per-row alignment is captured by
+///     `divisibility[innerDim]`. `innerStride = 1` implies stride-1
+///     contiguity; `innerStride = 0` implies inner-dim constancy.
+///
+/// Pessimistic / entry value: contiguity=1, constancy=1, divisibility=1,
+/// innerStride absent.
+struct AxisInfo {
+  SmallVector<int64_t> contiguity;
+  SmallVector<int64_t> constancy;
+  SmallVector<int64_t> divisibility;
+  std::optional<int64_t> knownConstant;
+  std::optional<int64_t> innerStride;
+
+  AxisInfo() = default;
+
+  static AxisInfo getPessimistic(unsigned rank) {
+    AxisInfo v;
+    v.contiguity.assign(rank, 1);
+    v.constancy.assign(rank, 1);
+    v.divisibility.assign(rank, 1);
+    return v;
+  }
+
+  unsigned getRank() const { return contiguity.size(); }
+  bool isInitialized() const { return getRank() > 0; }
+
+  bool operator==(const AxisInfo &rhs) const {
+    return contiguity == rhs.contiguity && constancy == rhs.constancy &&
+           divisibility == rhs.divisibility &&
+           knownConstant == rhs.knownConstant && innerStride == rhs.innerStride;
+  }
+
+  /// Conservative join. Two values reaching the same SSA value via different
+  /// control-flow paths must agree on what holds.
+  static AxisInfo join(const AxisInfo &lhs, const AxisInfo &rhs) {
+    if (!lhs.isInitialized())
+      return rhs;
+    if (!rhs.isInitialized())
+      return lhs;
+    assert(lhs.getRank() == rhs.getRank());
+    AxisInfo out;
+    unsigned r = lhs.getRank();
+    out.contiguity.resize(r);
+    out.constancy.resize(r);
+    out.divisibility.resize(r);
+    for (unsigned d = 0; d < r; ++d) {
+      out.contiguity[d] = std::min(lhs.contiguity[d], rhs.contiguity[d]);
+      out.constancy[d] = std::min(lhs.constancy[d], rhs.constancy[d]);
+      out.divisibility[d] = std::gcd(lhs.divisibility[d], rhs.divisibility[d]);
+    }
+    if (lhs.knownConstant && rhs.knownConstant &&
+        *lhs.knownConstant == *rhs.knownConstant)
+      out.knownConstant = lhs.knownConstant;
+    if (lhs.innerStride && rhs.innerStride &&
+        *lhs.innerStride == *rhs.innerStride)
+      out.innerStride = lhs.innerStride;
+    return out;
+  }
+
+  void print(raw_ostream &os) const {
+    os << "contiguity=[";
+    llvm::interleaveComma(contiguity, os);
+    os << "] constancy=[";
+    llvm::interleaveComma(constancy, os);
+    os << "] divisibility=[";
+    llvm::interleaveComma(divisibility, os);
+    os << "]";
+    if (knownConstant)
+      os << " const=" << *knownConstant;
+    if (innerStride)
+      os << " innerStride=" << *innerStride;
+  }
+};
+
+using AxisInfoLattice = dataflow::Lattice<AxisInfo>;
+
+/// Power-of-two divisor of `v`. Returns `kAxisInfoTop` when `v == 0`.
+static int64_t highestPow2Divisor(int64_t v) {
+  if (v == 0)
+    return kAxisInfoTop;
+  uint64_t u = static_cast<uint64_t>(std::abs(v));
+  return static_cast<int64_t>(u & (~u + 1));
+}
+
+/// Initial lattice value for an SSA value when no transfer function applies.
+static AxisInfo entryStateFor(Value v) {
+  if (auto vt = dyn_cast<VectorType>(v.getType()))
+    return AxisInfo::getPessimistic(vt.getRank());
+  return AxisInfo::getPessimistic(1);
+}
+
+/// AxisInfo for a tensor that is constant `c` everywhere.
+static AxisInfo splatAxisInfo(ArrayRef<int64_t> shape, int64_t c) {
+  AxisInfo v;
+  unsigned r = shape.size();
+  v.contiguity.assign(r, 1);
+  v.constancy.assign(shape.begin(), shape.end());
+  v.divisibility.assign(r, highestPow2Divisor(c));
+  v.knownConstant = c;
+  v.innerStride = 0;
+  return v;
+}
+
+/// Sparse forward dataflow analysis that computes `AxisInfo` for vector
+/// values reachable from the entry of the analyzed op.
+class AxisInfoAnalysis
+    : public dataflow::SparseForwardDataFlowAnalysis<AxisInfoLattice> {
+public:
+  MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(AxisInfoAnalysis)
+  using SparseForwardDataFlowAnalysis::SparseForwardDataFlowAnalysis;
+
+  LogicalResult visitOperation(Operation *op,
+                               ArrayRef<const AxisInfoLattice *> operands,
+                               ArrayRef<AxisInfoLattice *> results) override {
+    if (auto step = dyn_cast<vector::StepOp>(op))
+      return visitStep(step, results);
+    if (auto cst = dyn_cast<arith::ConstantOp>(op))
+      return visitConstant(cst, results);
+    if (auto bcast = dyn_cast<vector::BroadcastOp>(op))
+      return visitBroadcast(bcast, operands, results);
+    if (auto sc = dyn_cast<vector::ShapeCastOp>(op))
+      return visitShapeCast(sc, operands, results);
+    if (auto tp = dyn_cast<vector::TransposeOp>(op))
+      return visitTranspose(tp, operands, results);
+    if (auto add = dyn_cast<arith::AddIOp>(op))
+      return visitAddSub</*IsSub=*/false>(add, operands, results);
+    if (auto sub = dyn_cast<arith::SubIOp>(op))
+      return visitAddSub</*IsSub=*/true>(sub, operands, results);
+    if (auto mul = dyn_cast<arith::MulIOp>(op))
+      return visitMul(mul, operands, results);
+    if (auto div = dyn_cast<arith::DivUIOp>(op))
+      return visitDivRem</*IsSigned=*/false, /*IsRem=*/false>(div, operands,
+                                                              results);
+    if (auto div = dyn_cast<arith::DivSIOp>(op))
+      return visitDivRem</*IsSigned=*/true, /*IsRem=*/false>(div, operands,
+                                                             results);
+    if (auto rem = dyn_cast<arith::RemUIOp>(op))
+      return visitDivRem</*IsSigned=*/false, /*IsRem=*/true>(rem, operands,
+                                                             results);
+    if (auto rem = dyn_cast<arith::RemSIOp>(op))
+      return visitDivRem</*IsSigned=*/true, /*IsRem=*/true>(rem, operands,
+                                                            results);
+    if (auto andi = dyn_cast<arith::AndIOp>(op))
+      return visitAndI(andi, operands, results);
+    if (auto shl = dyn_cast<arith::ShLIOp>(op))
+      return visitShift</*IsLeft=*/true>(shl, operands, results);
+    if (auto shr = dyn_cast<arith::ShRUIOp>(op))
+      return visitShift</*IsLeft=*/false>(shr, operands, results);
+    if (auto sel = dyn_cast<arith::SelectOp>(op))
+      return visitSelect(sel, operands, results);
+    if (auto cast = dyn_cast<arith::IndexCastOp>(op))
+      return visitPassThrough(cast, operands, results);
+    if (auto cast = dyn_cast<arith::IndexCastUIOp>(op))
+      return visitPassThrough(cast, operands, results);
+    setAllPessimistic(op, results);
+    return success();
+  }
+
+  void setToEntryState(AxisInfoLattice *lattice) override {
+    propagateIfChanged(lattice,
+                       lattice->join(entryStateFor(lattice->getAnchor())));
+  }
+
+private:
+  void setAllPessimistic(Operation *op, ArrayRef<AxisInfoLattice *> results) {
+    for (auto [r, lat] : llvm::zip(op->getResults(), results)) {
+      AxisInfo state = entryStateFor(r);
+      propagateIfChanged(lat, lat->join(state));
+    }
+  }
+
+  // vector.step is always 1-D and produces [0, 1, ..., n-1].
+  LogicalResult visitStep(vector::StepOp op,
+                          ArrayRef<AxisInfoLattice *> results) {
+    auto vt = cast<VectorType>(op.getType());
+    int64_t n = vt.getNumElements();
+    AxisInfo v;
+    v.contiguity = {n};
+    v.constancy = {1};
+    v.divisibility = {kAxisInfoTop};
+    v.innerStride = 1;
+    propagateIfChanged(results[0], results[0]->join(v));
+    return success();
+  }
+
+  LogicalResult visitConstant(arith::ConstantOp op,
+                              ArrayRef<AxisInfoLattice *> results) {
+    auto vt = dyn_cast<VectorType>(op.getType());
+    if (!vt) {
+      if (auto intAttr = dyn_cast<IntegerAttr>(op.getValue())) {
+        int64_t c = intAttr.getValue().getSExtValue();
+        AxisInfo v;
+        v.contiguity = {1};
+        v.constancy = {1};
+        v.divisibility = {highestPow2Divisor(c)};
+        v.knownConstant = c;
+        propagateIfChanged(results[0], results[0]->join(v));
+        return success();
+      }
+      setAllPessimistic(op, results);
+      return success();
+    }
+    auto dense = dyn_cast<DenseIntElementsAttr>(op.getValue());
+    if (!dense) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    auto shape = vt.getShape();
+    if (dense.isSplat()) {
+      int64_t c = dense.getSplatValue<APInt>().getSExtValue();
+      AxisInfo v = splatAxisInfo(shape, c);
+      propagateIfChanged(results[0], results[0]->join(v));
+      return success();
+    }
+
+    // Compute innermost-dim contiguity / constancy / base-divisibility by
+    // iterating the dense values along the inner stride. Outer dims report
+    // pessimistic (1) unless they collapse trivially below.
+    unsigned r = shape.size();
+    int64_t inner = shape.back();
+    int64_t outer = vt.getNumElements() / inner;
+    if (inner < 2 || outer < 1) {
+      // Can't meaningfully analyze a 0/1-element inner dim; fall back to
+      // splat handling already covered, otherwise pessimistic.
+      AxisInfo v = AxisInfo::getPessimistic(r);
+      // For a 1-element inner dim the inner-dim contiguity/constancy is
+      // trivially 1 (already pessimistic).
+      propagateIfChanged(results[0], results[0]->join(v));
+      return success();
+    }
+    auto values = llvm::to_vector(dense.getValues<APInt>());
+    int64_t innerCont = inner;
+    int64_t innerConst = inner;
+    int64_t innerStride = values[1].getSExtValue() - values[0].getSExtValue();
+    int64_t base = values[0].getSExtValue();
+    int64_t baseDiv = highestPow2Divisor(base);
+    for (int64_t o = 0; o < outer; ++o) {
+      int64_t origin = values[o * inner].getSExtValue();
+      baseDiv = std::gcd(baseDiv, highestPow2Divisor(origin));
+      for (int64_t i = 1; i < inner; ++i) {
+        int64_t cur = values[o * inner + i].getSExtValue();
+        int64_t prev = values[o * inner + i - 1].getSExtValue();
+        int64_t diff = cur - prev;
+        if (diff != innerStride)
+          innerStride = std::numeric_limits<int64_t>::min(); // not AP
+        if (diff != 1)
+          innerCont = std::min<int64_t>(innerCont, i);
+        if (diff != 0)
+          innerConst = std::min<int64_t>(innerConst, i);
+      }
+    }
+    AxisInfo v = AxisInfo::getPessimistic(r);
+    if (innerStride == 1)
+      v.contiguity[r - 1] = innerCont;
+    else if (innerStride == 0)
+      v.constancy[r - 1] = innerConst;
+    // For a non-AP inner dim, leave at pessimistic.
+    v.divisibility[r - 1] = baseDiv;
+    if (innerStride != std::numeric_limits<int64_t>::min())
+      v.innerStride = innerStride;
+    propagateIfChanged(results[0], results[0]->join(v));
+    return success();
+  }
+
+  // vector.broadcast: source lattice extends to the broadcast dims with
+  // constancy = full extent on those dims. The trailing dims of the source
+  // (if any) align with the trailing dims of the result.
+  LogicalResult visitBroadcast(vector::BroadcastOp op,
+                               ArrayRef<const AxisInfoLattice *> operands,
+                               ArrayRef<AxisInfoLattice *> results) {
+    auto resTy = dyn_cast<VectorType>(op.getType());
+    if (!resTy) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    unsigned rRank = resTy.getRank();
+    AxisInfo src = operands[0]->getValue();
+    AxisInfo v = AxisInfo::getPessimistic(rRank);
+    auto resShape = resTy.getShape();
+    auto srcVt = dyn_cast<VectorType>(op.getSource().getType());
+    unsigned sRank = srcVt ? srcVt.getRank() : 0;
+    // Broadcast aligns trailing dims of the source with trailing dims of
+    // the result. Leading dims that are 1 in source (or absent) are filled
+    // with constancy = result extent.
+    for (unsigned d = 0; d < rRank; ++d) {
+      int64_t resExt = resShape[d];
+      // Index in source aligned with result dim d, or -1 if d is a
+      // broadcast (front-padded) dim.
+      int sIdx = static_cast<int>(d) - static_cast<int>(rRank - sRank);
+      if (sIdx < 0) {
+        v.constancy[d] = resExt;
+        v.contiguity[d] = 1;
+        v.divisibility[d] = src.isInitialized() ? src.divisibility.front() : 1;
+        continue;
+      }
+      int64_t srcExt = srcVt.getShape()[sIdx];
+      if (srcExt == 1 && resExt > 1) {
+        v.constancy[d] = resExt;
+        v.contiguity[d] = 1;
+        v.divisibility[d] = src.isInitialized() ? src.divisibility[sIdx] : 1;
+      } else if (src.isInitialized()) {
+        v.contiguity[d] = src.contiguity[sIdx];
+        v.constancy[d] = src.constancy[sIdx];
+        v.divisibility[d] = src.divisibility[sIdx];
+      }
+    }
+    if (src.knownConstant)
+      v.knownConstant = src.knownConstant;
+    // A broadcast that fans out a scalar / leading-1 source has the broadcast
+    // dim repeating its value -> inner stride 0. Otherwise, the trailing
+    // source dim's stride is preserved when its extent matches the result.
+    auto resShapeArr = resTy.getShape();
+    int64_t innerExt = resShapeArr.back();
+    int sIdxInner =
+        static_cast<int>(rRank - 1) - static_cast<int>(rRank - sRank);
+    if (sIdxInner < 0) {
+      v.innerStride = 0;
+    } else if (srcVt) {
+      int64_t srcInner = srcVt.getShape()[sIdxInner];
+      if (srcInner == 1 && innerExt > 1)
+        v.innerStride = 0;
+      else if (srcInner == innerExt && src.innerStride)
+        v.innerStride = src.innerStride;
+    }
+    propagateIfChanged(results[0], results[0]->join(v));
+    return success();
+  }
+
+  // vector.shape_cast: handle two cases.
+  //   (a) Identity-like: shapes match after stripping leading-1 dims —
+  //       rebind per-dim info to the new dim positions.
+  //   (b) General reshape with the same total element count and row-major
+  //       linearization — propagate the source's innermost-dim info (inner
+  //       contiguity / constancy) to the destination's innermost dim,
+  //       capped by the inner extent. Outer dims stay pessimistic.
+  LogicalResult visitShapeCast(vector::ShapeCastOp op,
+                               ArrayRef<const AxisInfoLattice *> operands,
+                               ArrayRef<AxisInfoLattice *> results) {
+    auto srcTy = dyn_cast<VectorType>(op.getSource().getType());
+    auto dstTy = dyn_cast<VectorType>(op.getType());
+    if (!srcTy || !dstTy) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    AxisInfo src = operands[0]->getValue();
+    if (!src.isInitialized()) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    // Strip leading 1-dims on both sides; if remaining shapes match, this
+    // is an identity-like reshape.
+    auto stripLeading = [](ArrayRef<int64_t> s) {
+      unsigned i = 0;
+      while (i < s.size() && s[i] == 1)
+        ++i;
+      return s.drop_front(i);
+    };
+    auto sCore = stripLeading(srcTy.getShape());
+    auto dCore = stripLeading(dstTy.getShape());
+    unsigned dRank = dstTy.getRank();
+    AxisInfo v = AxisInfo::getPessimistic(dRank);
+    if (sCore == dCore) {
+      unsigned sLead = srcTy.getRank() - sCore.size();
+      unsigned dLead = dRank - dCore.size();
+      for (unsigned d = dLead; d < dRank; ++d) {
+        unsigned sIdx = sLead + (d - dLead);
+        v.contiguity[d] = src.contiguity[sIdx];
+        v.constancy[d] = src.constancy[sIdx];
+        v.divisibility[d] = src.divisibility[sIdx];
+      }
+    } else {
+      // General linear reshape. Propagate source's inner-dim contiguity /
+      // constancy to dst's inner dim, capped by inner extent. Treat inner
+      // info conservatively as the min across all source dims (so a 1-D
+      // source with full contig => inner-dim contig on dst; an N-D source
+      // collapsed to 1-D inherits the inner-dim info).
+      int64_t innerExt = dstTy.getShape().back();
+      int64_t srcContig = std::numeric_limits<int64_t>::max();
+      int64_t srcConst = std::numeric_limits<int64_t>::max();
+      int64_t srcDiv = src.divisibility[src.getRank() - 1];
+      for (unsigned d = 0; d < src.getRank(); ++d) {
+        srcContig = std::min(srcContig, src.contiguity[d]);
+        srcConst = std::min(srcConst, src.constancy[d]);
+      }
+      v.contiguity[dRank - 1] = std::min<int64_t>(srcContig, innerExt);
+      v.constancy[dRank - 1] = std::min<int64_t>(srcConst, innerExt);
+      v.divisibility[dRank - 1] = srcDiv;
+    }
+    if (src.knownConstant)
+      v.knownConstant = src.knownConstant;
+    // Identity-like and general row-major reshape both preserve the source
+    // inner-stride property when the source has a single AP characterization.
+    if (src.innerStride)
+      v.innerStride = src.innerStride;
+    propagateIfChanged(results[0], results[0]->join(v));
+    return success();
+  }
+
+  // vector.transpose: permute per-dim contiguity / constancy / divisibility
+  // according to the transpose permutation. permutation[i] is the source
+  // dim that ends up at result dim i.
+  LogicalResult visitTranspose(vector::TransposeOp op,
+                               ArrayRef<const AxisInfoLattice *> operands,
+                               ArrayRef<AxisInfoLattice *> results) {
+    auto resTy = dyn_cast<VectorType>(op.getType());
+    if (!resTy) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    AxisInfo src = operands[0]->getValue();
+    if (!src.isInitialized()) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    ArrayRef<int64_t> perm = op.getPermutation();
+    unsigned r = resTy.getRank();
+    AxisInfo v = AxisInfo::getPessimistic(r);
+    for (unsigned d = 0; d < r; ++d) {
+      unsigned s = static_cast<unsigned>(perm[d]);
+      v.contiguity[d] = src.contiguity[s];
+      v.constancy[d] = src.constancy[s];
+      v.divisibility[d] = src.divisibility[s];
+    }
+    if (src.knownConstant)
+      v.knownConstant = src.knownConstant;
+    // innerStride only survives when the new inner dim came from the old
+    // inner dim (otherwise a different axis is now the contiguous one).
+    if (src.innerStride && perm.back() == src.getRank() - 1)
+      v.innerStride = src.innerStride;
+    propagateIfChanged(results[0], results[0]->join(v));
+    return success();
+  }
+
+  template <bool IsSub, typename OpTy>
+  LogicalResult visitAddSub(OpTy op, ArrayRef<const AxisInfoLattice *> operands,
+                            ArrayRef<AxisInfoLattice *> results) {
+    auto vt = dyn_cast<VectorType>(op.getType());
+    if (!vt) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    AxisInfo lhs = operands[0]->getValue();
+    AxisInfo rhs = operands[1]->getValue();
+    if (!lhs.isInitialized() || !rhs.isInitialized()) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    unsigned r = vt.getRank();
+    AxisInfo v = AxisInfo::getPessimistic(r);
+    for (unsigned d = 0; d < r; ++d) {
+      int64_t lhsCont = lhs.contiguity[d];
+      int64_t rhsCont = rhs.contiguity[d];
+      int64_t lhsConst = lhs.constancy[d];
+      int64_t rhsConst = rhs.constancy[d];
+      // contiguity propagates through add when one side is constant on the
+      // run, and through sub only when the rhs is constant on the run.
+      int64_t cont = IsSub ? std::min(lhsCont, rhsConst)
+                           : std::max(std::min(lhsCont, rhsConst),
+                                      std::min(rhsCont, lhsConst));
+      v.contiguity[d] = std::max<int64_t>(1, cont);
+      v.constancy[d] = std::min(lhsConst, rhsConst);
+      v.divisibility[d] = std::gcd(lhs.divisibility[d], rhs.divisibility[d]);
+    }
+    // x + uniform-c: stride preserved. x - uniform-c: same. uniform-c - x:
+    // stride flips sign (only useful for the "stride 0" case, which it
+    // preserves trivially).
+    auto isUniform = [&](const AxisInfo &a) {
+      unsigned inner = vt.getRank() - 1;
+      return a.constancy[inner] >= vt.getShape()[inner];
+    };
+    if (lhs.innerStride && isUniform(rhs)) {
+      v.innerStride = *lhs.innerStride;
+    } else if (rhs.innerStride && isUniform(lhs)) {
+      v.innerStride = IsSub ? -*rhs.innerStride : *rhs.innerStride;
+    }
+    propagateIfChanged(results[0], results[0]->join(v));
+    return success();
+  }
+
+  LogicalResult visitMul(arith::MulIOp op,
+                         ArrayRef<const AxisInfoLattice *> operands,
+                         ArrayRef<AxisInfoLattice *> results) {
+    auto vt = dyn_cast<VectorType>(op.getType());
+    if (!vt) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    AxisInfo lhs = operands[0]->getValue();
+    AxisInfo rhs = operands[1]->getValue();
+    if (!lhs.isInitialized() || !rhs.isInitialized()) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    unsigned r = vt.getRank();
+    auto shape = vt.getShape();
+    AxisInfo v = AxisInfo::getPessimistic(r);
+    auto unitConstant = [](const AxisInfo &a, unsigned d, int64_t lanes) {
+      return a.knownConstant && *a.knownConstant == 1 &&
+             a.constancy[d] >= lanes;
+    };
+    for (unsigned d = 0; d < r; ++d) {
+      v.constancy[d] = std::min({shape[d], lhs.constancy[d], rhs.constancy[d]});
+      v.divisibility[d] = std::min<int64_t>(
+          kAxisInfoTop, lhs.divisibility[d] * rhs.divisibility[d]);
+      // Multiplying by uniform `s` only keeps contiguity when `s == 1`.
+      if (unitConstant(lhs, d, shape[d]))
+        v.contiguity[d] = std::min(rhs.contiguity[d], shape[d]);
+      else if (unitConstant(rhs, d, shape[d]))
+        v.contiguity[d] = std::min(lhs.contiguity[d], shape[d]);
+      else
+        v.contiguity[d] = 1;
+    }
+    // x * uniform-c: stride scales by c. (Both operands uniform => 0.)
+    unsigned inner = vt.getRank() - 1;
+    auto isUniformInner = [&](const AxisInfo &a) {
+      return a.constancy[inner] >= shape[inner];
+    };
+    if (lhs.innerStride && isUniformInner(rhs) && rhs.knownConstant) {
+      v.innerStride = *lhs.innerStride * *rhs.knownConstant;
+    } else if (rhs.innerStride && isUniformInner(lhs) && lhs.knownConstant) {
+      v.innerStride = *rhs.innerStride * *lhs.knownConstant;
+    }
+    propagateIfChanged(results[0], results[0]->join(v));
+    return success();
+  }
+
+  // arith.divui / arith.divsi / arith.remui / arith.remsi by a uniform
+  // positive constant `c`.
+  //
+  // Division (IsRem=false): when the lhs is an inner-dim AP `(base, s, n)`
+  // with `c | s` and `c | divisibility[inner]` (so the per-row base / c is
+  // exact), the result is an AP with stride `s / c` and inner divisibility
+  // `divisibility[inner] / c`. Special cases: `s/c == 1` flags inner-dim
+  // contiguity; `s/c == 0` flags inner-dim constancy.
+  //
+  // Remainder (IsRem=true): when `c | s`, every element of a row sits at
+  // the same residue class -> inner-dim constant -> `innerStride = 0`,
+  // `constancy[inner] = inner` (matches the analysis's notion of
+  // chunk-uniform values along the inner dim).
+  //
+  // Signed vs unsigned only differs in the constant interpretation; we
+  // require positive constants so the signed/unsigned distinction is moot
+  // here.
+  template <bool IsSigned, bool IsRem, typename OpTy>
+  LogicalResult visitDivRem(OpTy op, ArrayRef<const AxisInfoLattice *> operands,
+                            ArrayRef<AxisInfoLattice *> results) {
+    auto vt = dyn_cast<VectorType>(op.getType());
+    if (!vt) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    AxisInfo lhs = operands[0]->getValue();
+    AxisInfo rhs = operands[1]->getValue();
+    if (!lhs.isInitialized() || !rhs.isInitialized()) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    unsigned r = vt.getRank();
+    unsigned inner = r - 1;
+    auto shape = vt.getShape();
+    AxisInfo v = AxisInfo::getPessimistic(r);
+
+    bool rhsUniform = rhs.constancy[inner] >= shape[inner] && rhs.knownConstant;
+    if (!rhsUniform || *rhs.knownConstant <= 0) {
+      propagateIfChanged(results[0], results[0]->join(v));
+      return success();
+    }
+    int64_t c = *rhs.knownConstant;
+
+    if (!lhs.innerStride) {
+      propagateIfChanged(results[0], results[0]->join(v));
+      return success();
+    }
+    int64_t s = *lhs.innerStride;
+    int64_t baseDivLhs = lhs.divisibility[inner];
+    if (s % c != 0) {
+      propagateIfChanged(results[0], results[0]->join(v));
+      return success();
+    }
+
+    if (IsRem) {
+      // (base + i*s) mod c, with c | s, is the constant base mod c.
+      v.innerStride = 0;
+      v.constancy[inner] = shape[inner];
+      // The remainder is in [0, c-1], so any power-of-two divisor of c is a
+      // lower bound on alignment. Use lhs's existing divisibility too.
+      v.divisibility[inner] = std::gcd(baseDivLhs, highestPow2Divisor(c));
+    } else {
+      if (baseDivLhs % c != 0) {
+        propagateIfChanged(results[0], results[0]->join(v));
+        return success();
+      }
+      int64_t newStride = s / c;
+      v.innerStride = newStride;
+      if (newStride == 1)
+        v.contiguity[inner] = shape[inner];
+      else if (newStride == 0)
+        v.constancy[inner] = shape[inner];
+      v.divisibility[inner] = baseDivLhs / c;
+    }
+    propagateIfChanged(results[0], results[0]->join(v));
+    return success();
+  }
+
+  // arith.andi: `x & m` with a uniform positive constant mask `m`.
+  // The most useful case is `x & (P - 1)` for `P` a power of 2: this is
+  // equivalent to `x mod P`, so when the lhs is an inner-dim AP with stride
+  // divisible by `P` the result is inner-dim constant. We also handle the
+  // trivial `m == 0` (always zero) and `m == -1`/all-ones (identity)
+  // shapes.
+  LogicalResult visitAndI(arith::AndIOp op,
+                          ArrayRef<const AxisInfoLattice *> operands,
+                          ArrayRef<AxisInfoLattice *> results) {
+    auto vt = dyn_cast<VectorType>(op.getType());
+    if (!vt) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    AxisInfo lhs = operands[0]->getValue();
+    AxisInfo rhs = operands[1]->getValue();
+    if (!lhs.isInitialized() || !rhs.isInitialized()) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    unsigned r = vt.getRank();
+    unsigned inner = r - 1;
+    auto shape = vt.getShape();
+    AxisInfo v = AxisInfo::getPessimistic(r);
+
+    // Look for a uniform constant mask on either side.
+    auto getUniformMask = [&](const AxisInfo &a) -> std::optional<int64_t> {
+      if (a.constancy[inner] >= shape[inner] && a.knownConstant)
+        return a.knownConstant;
+      return std::nullopt;
+    };
+    std::optional<int64_t> mLhs = getUniformMask(lhs);
+    std::optional<int64_t> mRhs = getUniformMask(rhs);
+    if (!mLhs && !mRhs) {
+      propagateIfChanged(results[0], results[0]->join(v));
+      return success();
+    }
+    const AxisInfo &x = mLhs ? rhs : lhs;
+    int64_t m = mLhs ? *mLhs : *mRhs;
+
+    if (m == 0) {
+      v.knownConstant = 0;
+      v.innerStride = 0;
+      v.constancy[inner] = shape[inner];
+      v.divisibility[inner] = kAxisInfoTop;
+      propagateIfChanged(results[0], results[0]->join(v));
+      return success();
+    }
+
+    // `m == P - 1` with P a power of 2 -> equivalent to `x mod P`.
+    if (m > 0 && llvm::isPowerOf2_64(static_cast<uint64_t>(m + 1))) {
+      int64_t P = m + 1;
+      if (x.innerStride && *x.innerStride % P == 0) {
+        v.innerStride = 0;
+        v.constancy[inner] = shape[inner];
+        v.divisibility[inner] =
+            std::gcd(x.divisibility[inner], highestPow2Divisor(P));
+        propagateIfChanged(results[0], results[0]->join(v));
+        return success();
+      }
+    }
+    // Conservative fallback for unrecognized masks.
+    propagateIfChanged(results[0], results[0]->join(v));
+    return success();
+  }
+
+  // arith.shli (left shift) / arith.shrui (logical right shift) by a
+  // uniform constant `k`. These are `* (1 << k)` and `/ (1 << k)`
+  // (truncating, but for non-negative values the trunc is exact when
+  // `(1 << k)` divides the value). We model them by reducing to mul/divui.
+  template <bool IsLeft, typename OpTy>
+  LogicalResult visitShift(OpTy op, ArrayRef<const AxisInfoLattice *> operands,
+                           ArrayRef<AxisInfoLattice *> results) {
+    auto vt = dyn_cast<VectorType>(op.getType());
+    if (!vt) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    AxisInfo lhs = operands[0]->getValue();
+    AxisInfo rhs = operands[1]->getValue();
+    if (!lhs.isInitialized() || !rhs.isInitialized()) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    unsigned r = vt.getRank();
+    unsigned inner = r - 1;
+    auto shape = vt.getShape();
+    AxisInfo v = AxisInfo::getPessimistic(r);
+
+    if (rhs.constancy[inner] < shape[inner] || !rhs.knownConstant) {
+      propagateIfChanged(results[0], results[0]->join(v));
+      return success();
+    }
+    int64_t k = *rhs.knownConstant;
+    if (k < 0 || k >= 63) {
+      propagateIfChanged(results[0], results[0]->join(v));
+      return success();
+    }
+    int64_t factor = 1LL << k;
+
+    if (IsLeft) {
+      // x << k == x * factor.
+      if (lhs.innerStride) {
+        v.innerStride = *lhs.innerStride * factor;
+        if (*v.innerStride == 1)
+          v.contiguity[inner] = shape[inner];
+        else if (*v.innerStride == 0)
+          v.constancy[inner] = shape[inner];
+      }
+      v.divisibility[inner] =
+          std::min<int64_t>(kAxisInfoTop, lhs.divisibility[inner] * factor);
+    } else {
+      // x >> k == x / factor (for non-negative x); same conditions as divui.
+      if (lhs.innerStride && *lhs.innerStride % factor == 0 &&
+          lhs.divisibility[inner] % factor == 0) {
+        int64_t newStride = *lhs.innerStride / factor;
+        v.innerStride = newStride;
+        if (newStride == 1)
+          v.contiguity[inner] = shape[inner];
+        else if (newStride == 0)
+          v.constancy[inner] = shape[inner];
+        v.divisibility[inner] = lhs.divisibility[inner] / factor;
+      }
+    }
+    propagateIfChanged(results[0], results[0]->join(v));
+    return success();
+  }
+
+  // arith.select: result is at least as constrained as the meet of the two
+  // arms. We propagate fields where both arms agree.
+  LogicalResult visitSelect(arith::SelectOp op,
+                            ArrayRef<const AxisInfoLattice *> operands,
+                            ArrayRef<AxisInfoLattice *> results) {
+    auto vt = dyn_cast<VectorType>(op.getType());
+    if (!vt) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    // operands: [cond, true, false]
+    AxisInfo t = operands[1]->getValue();
+    AxisInfo f = operands[2]->getValue();
+    if (!t.isInitialized() || !f.isInitialized()) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    AxisInfo v = AxisInfo::join(t, f);
+    propagateIfChanged(results[0], results[0]->join(v));
+    return success();
+  }
+
+  template <typename OpTy>
+  LogicalResult visitPassThrough(OpTy op,
+                                 ArrayRef<const AxisInfoLattice *> operands,
+                                 ArrayRef<AxisInfoLattice *> results) {
+    if (!isa<VectorType>(op.getType())) {
+      setAllPessimistic(op, results);
+      return success();
+    }
+    propagateIfChanged(results[0], results[0]->join(operands[0]->getValue()));
+    return success();
+  }
+};
+
+} // namespace mlir::xegpu::detail::axis_dataflow
+
+namespace {
+
+using ::mlir::xegpu::detail::axis_dataflow::AxisInfo;
+using ::mlir::xegpu::detail::axis_dataflow::AxisInfoLattice;
+
+//===----------------------------------------------------------------------===//
+// Coalescing decision.
+//===----------------------------------------------------------------------===//
+
+struct CoalesceDecision {
+  enum class Kind { None, Chunked };
+  Kind kind = Kind::None;
+  int64_t laneLayout = 1; // lane_layout along the innermost dim
+  int64_t factor = 1;     // lane_data factor along the innermost dim
+};
+
+/// Largest power-of-two `<= bound` that divides `numLanes`.
+static int64_t largestPow2Divisor(int64_t numLanes, int64_t bound) {
+  if (bound < 2 || numLanes < 2)
+    return 1;
+  int64_t f = std::min<int64_t>(bound, numLanes);
+  // Round down to power of 2.
+  if (!llvm::isPowerOf2_64(f))
+    f = static_cast<int64_t>(llvm::bit_floor(static_cast<uint64_t>(f)));
+  while (f >= 2) {
+    if (numLanes % f == 0)
+      return f;
+    f /= 2;
+  }
+  return 1;
+}
+
+/// Decide how to coalesce given the offsets axis info.
+///
+/// We pick `lane_layout[inner]` first using the same default rule as
+/// `XeGPUPropagateLayout`: `lane_layout[inner] = min(subgroupSize, inner)`,
+/// rounded down to a power-of-2 divisor of `inner`. The remaining lane
+/// budget then becomes `lane_data[inner] = inner / lane_layout`, capped by
+/// the per-lane chunk-size budget and the offsets contiguity.
+static CoalesceDecision decide(const AxisInfo &info,
+                               ArrayRef<int64_t> offsetsShape,
+                               int64_t origChunk, unsigned maxChunkSize,
+                               unsigned subgroupSize) {
+  CoalesceDecision d;
+  if (!info.isInitialized() || offsetsShape.empty())
+    return d;
+  unsigned innerDim = offsetsShape.size() - 1;
+  int64_t inner = offsetsShape[innerDim];
+  if (inner < 2)
+    return d;
+
+  // All-equal offsets: every lane sees the same address. There is no
+  // layout-only encoding for this; we leave the op alone.
+  if (info.constancy[innerDim] >= inner)
+    return d;
+
+  // Pick lane_layout first: PropagateLayout's default for a 1-D / inner dim
+  // is `min(subgroupSize, inner)`, rounded down to a divisor of inner.
+  int64_t laneLayout =
+      largestPow2Divisor(inner, std::min<int64_t>(subgroupSize, inner));
+  if (laneLayout < 1)
+    laneLayout = 1;
+
+  // Each lane sees `inner / laneLayout` elements of the offsets vector.
+  int64_t perLane = inner / laneLayout;
+  if (perLane < 2)
+    return d; // already one element per lane, nothing to coalesce.
+
+  if (origChunk < 1)
+    origChunk = 1;
+  int64_t budget = static_cast<int64_t>(maxChunkSize) / origChunk;
+  if (budget < 2)
+    return d;
+
+  int64_t bound = std::min<int64_t>(info.contiguity[innerDim], budget);
+  bound = std::min<int64_t>(bound, perLane);
+  if (bound < 2)
+    return d;
+
+  int64_t factor = largestPow2Divisor(perLane, bound);
+  if (factor < 2)
+    return d;
+  d.kind = CoalesceDecision::Kind::Chunked;
+  d.laneLayout = laneLayout;
+  d.factor = factor;
+  return d;
+}
+
+/// Returns true if `mask` is a constant `dense<true>` vector.
+static bool isAllTrueMask(Value mask) {
+  auto vecTy = dyn_cast<VectorType>(mask.getType());
+  if (!vecTy)
+    return false;
+  auto cst = mask.getDefiningOp<arith::ConstantOp>();
+  if (!cst)
+    return false;
+  auto dense = dyn_cast<DenseIntElementsAttr>(cst.getValue());
+  if (!dense || !dense.isSplat())
+    return false;
+  return dense.getSplatValue<APInt>().getBoolValue();
+}
+
+/// Build a `lane_layout`/`lane_data`/`inst_data` layout of rank `rank`,
+/// with the given lane_layout / lane_data on the innermost dim (1
+/// elsewhere). `inst_data` is `lane_layout * lane_data` per dim, so the
+/// invariant `inst_data[d] == lane_layout[d] * lane_data[d]` holds.
+static xegpu::LayoutAttr buildLaneDataLayout(MLIRContext *ctx, unsigned rank,
+                                             int64_t innerLaneLayout,
+                                             int64_t innerLaneData) {
+  SmallVector<int32_t> laneLayout(rank, 1);
+  SmallVector<int32_t> laneData(rank, 1);
+  SmallVector<int32_t> instData(rank, 1);
+  laneLayout.back() = static_cast<int32_t>(innerLaneLayout);
+  laneData.back() = static_cast<int32_t>(innerLaneData);
+  instData.back() = static_cast<int32_t>(innerLaneLayout * innerLaneData);
+  return xegpu::LayoutAttr::get(ctx, instData, laneLayout, laneData);
+}
+
+//===----------------------------------------------------------------------===//
+// Rewrites.
+//===----------------------------------------------------------------------===//
+
+namespace {
+
+/// Look up the subgroup size from the enclosing gpu.module's xevm.target.
+/// Falls back to 16 when no target chip is found or the chip is unknown,
+/// matching the typical Intel Xe2 default. This keeps the pass usable on
+/// plain `module { ... }` IR (e.g. unit lit tests) where there's no
+/// gpu.module / xevm.target wrapper.
+static unsigned lookupSubgroupSize(Operation *op) {
+  const auto *uArch =
+      xegpu::uArch::getUArch(xegpu::getChipStr(op).value_or(""));
+  return uArch ? static_cast<unsigned>(uArch->getSubgroupSize()) : 16u;
+}
+
+/// Common analysis preconditions: vector offsets/value, all-true mask,
+/// no existing non-trivial lane_data, no explicit chunk_size > 1.
+template <typename OpTy>
+static bool isCandidateForCoalesce(OpTy op) {
+  auto offsetsTy = dyn_cast<VectorType>(op.getOffsets().getType());
+  if (!offsetsTy || offsetsTy.getNumElements() <= 1)
+    return false;
+  if (!op.getValueType())
+    return false;
+  if (!isAllTrueMask(op.getMask()))
+    return false;
+  if (auto layout = op.getLayoutAttr())
+    if (!layout.getEffectiveLaneDataAsInt().empty())
+      return false;
+  if (op.getChunkSizeAttr() && op.getChunkSize().value_or(1) > 1)
+    return false;
+  return true;
+}
+
+/// True when `op` (a gather load or scatter store) is tied to a
+/// `vector.multi_reduction`: a load whose result feeds a reduction, or a
+/// store whose stored value comes from one (through layout-neutral /
+/// elementwise / insert glue).
+///
+/// Coalescing such an access is gated off for now: the analysis only sets
+/// `lane_data[FCD]`, and consuming that on a reduction requires reduction-
+/// aware layout handling that is not part of this change. A coalesced store
+/// fed (transitively) by a reduction would seed `lane_data[FCD] = N` that the
+/// reduction result (kept at `lane_data = 1`) cannot match, producing an
+/// unlowerable `xegpu.convert_layout lane_data=[1]<->[N]`. Follow-up PRs that
+/// add reduction coalescing relax this gate.
+template <typename OpTy>
+static bool isReductionTied(OpTy op) {
+  if constexpr (std::is_same_v<OpTy, xegpu::StoreScatterOp>) {
+    // Backward walk of the stored value's slice through the reassembly
+    // (convert_layout / shape_cast / insert(_strided_slice) / elementwise)
+    // that typically sits between a reduction and the store.
+    SmallVector<Value, 8> worklist{op.getValue()};
+    llvm::SmallPtrSet<Operation *, 16> seen;
+    unsigned steps = 0;
+    while (!worklist.empty() && steps++ < 64) {
+      Value v = worklist.pop_back_val();
+      Operation *def = v.getDefiningOp();
+      if (!def || !seen.insert(def).second)
+        continue;
+      if (isa<vector::MultiDimReductionOp>(def))
+        return true;
+      if (isa<vector::ShapeCastOp, vector::BitCastOp, xegpu::ConvertLayoutOp,
+              vector::InsertOp, vector::InsertStridedSliceOp>(def) ||
+          OpTrait::hasElementwiseMappableTraits(def))
+        for (Value operand : def->getOperands())
+          if (isa<VectorType>(operand.getType()))
+            worklist.push_back(operand);
+    }
+    return false;
+  } else {
+    for (Operation *user : op->getUsers()) {
+      Operation *u = user;
+      while (
+          u &&
+          isa<vector::ShapeCastOp, vector::BitCastOp, xegpu::ConvertLayoutOp>(
+              u)) {
+        if (u->getNumResults() != 1 || u->getResult(0).use_empty())
+          break;
+        u = *u->getResult(0).getUsers().begin();
+      }
+      if (u && isa<vector::MultiDimReductionOp>(u))
+        return true;
+    }
+    return false;
+  }
+}
+
+/// Run the analysis on a single op. If the offsets analyze as `Chunked`,
+/// stamp a `xegpu.coalesce_hint` attribute carrying the FCD lane_data
+/// factor.
+template <typename OpTy>
+static void analyzeAndStampHint(OpTy op, DataFlowSolver &solver,
+                                unsigned maxChunkSize) {
+  if (!isCandidateForCoalesce(op))
+    return;
+  // Do not coalesce accesses tied to a reduction (see isReductionTied);
+  // reduction coalescing is added in follow-up PRs.
+  if (isReductionTied(op))
+    return;
+  auto offsetsTy = cast<VectorType>(op.getOffsets().getType());
+  unsigned subgroupSize = lookupSubgroupSize(op);
+  const auto *lat = solver.lookupState<AxisInfoLattice>(op.getOffsets());
+  if (!lat || !lat->getValue().isInitialized())
+    return;
+  int64_t origChunk = static_cast<int64_t>(op.getChunkSize().value_or(1));
+  auto d = decide(lat->getValue(), offsetsTy.getShape(), origChunk,
+                  maxChunkSize, subgroupSize);
+  if (d.kind != CoalesceDecision::Kind::Chunked)
+    return;
+  auto hint = xegpu::CoalesceHintAttr::get(op.getContext(), d.factor);
+  op->setAttr(xegpu::getCoalesceHintAttrName(), hint);
+}
+
+/// Apply a stamped hint on `op`: build a lane_layout/lane_data/inst_data
+/// layout from the hint's `factor` and the op's offsets inner extent +
+/// chip-derived subgroup size, install it, drop a trivial `chunk_size = 1`,
+/// and remove the hint. Returns success on apply (or no-op when no hint),
+/// failure when the hint is malformed.
+template <typename OpTy>
+static LogicalResult applyHintOnOp(OpTy op) {
+  auto hint = op->template getAttrOfType<xegpu::CoalesceHintAttr>(
+      xegpu::getCoalesceHintAttrName());
+  if (!hint)
+    return success(); // no hint: idempotent no-op.
+
+  auto offsetsTy = dyn_cast<VectorType>(op.getOffsets().getType());
+  auto valueTy = op.getValueType();
+  if (!offsetsTy || !valueTy || offsetsTy.getNumElements() <= 1)
+    return failure();
+
+  int64_t factor = hint.getFactor().getInt();
+  int64_t inner = offsetsTy.getShape().back();
+  unsigned subgroupSize = lookupSubgroupSize(op);
+  int64_t laneLayout =
+      largestPow2Divisor(inner, std::min<int64_t>(subgroupSize, inner));
+  if (laneLayout < 1 || inner % (laneLayout * factor) != 0)
+    return failure();
+
+  auto layout = buildLaneDataLayout(op.getContext(), valueTy.getRank(),
+                                    laneLayout, factor);
+  int64_t origChunk = static_cast<int64_t>(op.getChunkSize().value_or(1));
+  bool dropChunk = op.getChunkSizeAttr() && origChunk == 1 && factor > 1;
+  op.setLayoutAttr(layout);
+  if (dropChunk)
+    op.removeChunkSizeAttr();
+  op->removeAttr(xegpu::getCoalesceHintAttrName());
+  return success();
+}
+
+} // namespace
+
+} // namespace
+
+//===----------------------------------------------------------------------===//
+// Public APIs.
+//===----------------------------------------------------------------------===//
+
+void mlir::xegpu::runCoalesceGatherScatterAnalysis(
+    Operation *root, const CoalesceGatherScatterAnalysisOptions &options) {
+  DataFlowSolver solver;
+  solver.load<dataflow::DeadCodeAnalysis>();
+  solver.load<mlir::xegpu::detail::axis_dataflow::AxisInfoAnalysis>();
+  if (failed(solver.initializeAndRun(root)))
+    return;
+
+  root->walk([&](Operation *op) {
+    if (auto load = dyn_cast<xegpu::LoadGatherOp>(op))
+      analyzeAndStampHint(load, solver, options.maxChunkSize);
+    else if (auto store = dyn_cast<xegpu::StoreScatterOp>(op))
+      analyzeAndStampHint(store, solver, options.maxChunkSize);
+  });
+}
+
+LogicalResult mlir::xegpu::applyCoalesceGatherScatterHint(Operation *op) {
+  if (auto load = dyn_cast<xegpu::LoadGatherOp>(op))
+    return applyHintOnOp(load);
+  if (auto store = dyn_cast<xegpu::StoreScatterOp>(op))
+    return applyHintOnOp(store);
+  // Hint attached to an unsupported op: silently drop it.
+  if (op->hasAttr(xegpu::getCoalesceHintAttrName()))
+    op->removeAttr(xegpu::getCoalesceHintAttrName());
+  return success();
+}
+
+void mlir::xegpu::applyCoalesceGatherScatterHints(Operation *root) {
+  root->walk([&](Operation *op) {
+    if (op->hasAttr(xegpu::getCoalesceHintAttrName()))
+      (void)applyCoalesceGatherScatterHint(op);
+  });
+}
+
+void mlir::xegpu::clearCoalesceGatherScatterHints(Operation *root) {
+  StringRef name = xegpu::getCoalesceHintAttrName();
+  root->walk([&](Operation *op) {
+    if (op->hasAttr(name))
+      op->removeAttr(name);
+  });
+}
diff --git a/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir b/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir
new file mode 100644
index 0000000000000..ddfcf6e506c94
--- /dev/null
+++ b/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir
@@ -0,0 +1,105 @@
+// RUN: mlir-opt -split-input-file \
+// RUN:   -test-xegpu-coalesce-gather-scatter="analyze-only=true" %s | FileCheck %s
+
+// Analyze-only mode: stamps `xegpu.coalesce_hint` on coalescible ops and
+// leaves the layout / chunk_size unchanged. This test pins the hint
+// attribute contract that the apply API (and downstream propagator
+// integrations) consume.
+
+// -----
+// 1-D vector.step, fully coalescible -> hint with factor = 2 stamped.
+// CHECK-LABEL: func.func @load_step_offsets(
+// CHECK: xegpu.load
+// CHECK-SAME: <{chunk_size = 1 : i64}>
+// CHECK-SAME: {xegpu.coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+// CHECK-NOT: lane_data
+func.func @load_step_offsets(%ptr: i64) -> vector<32xf32> {
+  %offsets = vector.step : vector<32xindex>
+  %mask = arith.constant dense<true> : vector<32xi1>
+  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+  return %v : vector<32xf32>
+}
+
+// -----
+// 2-D leading-1 dim: hint stamped on the load with factor = 2.
+// CHECK-LABEL: func.func @load_2d_leading_unit(
+// CHECK: xegpu.load
+// CHECK-SAME: {xegpu.coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}
+// CHECK-SAME: : i64, vector<1x32xindex>, vector<1x32xi1> -> vector<1x32xf32>
+// CHECK-NOT: lane_data
+func.func @load_2d_leading_unit(%ptr: i64) -> vector<1x32xf32> {
+  %step = vector.step : vector<32xindex>
+  %offsets = vector.shape_cast %step : vector<32xindex> to vector<1x32xindex>
+  %mask = arith.constant dense<true> : vector<1x32xi1>
+  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : i64, vector<1x32xindex>, vector<1x32xi1> -> vector<1x32xf32>
+  return %v : vector<1x32xf32>
+}
+
+// -----
+// Stride-4 offsets: not coalescible, no hint stamped.
+// CHECK-LABEL: func.func @load_stride4_no_hint(
+// CHECK: xegpu.load
+// CHECK-NOT: coalesce_hint
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+func.func @load_stride4_no_hint(%ptr: i64) -> vector<32xf32> {
+  %c4 = arith.constant 4 : index
+  %step = vector.step : vector<32xindex>
+  %splat = vector.broadcast %c4 : index to vector<32xindex>
+  %offsets = arith.muli %step, %splat : vector<32xindex>
+  %mask = arith.constant dense<true> : vector<32xi1>
+  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+  return %v : vector<32xf32>
+}
+
+// -----
+// All-equal offsets: classified as Broadcast by decide(); analysis returns
+// None at the stamping stage, so no hint is stamped (broadcast-load
+// rewrite was removed).
+// CHECK-LABEL: func.func @load_broadcast_offsets_no_hint(
+// CHECK: xegpu.load
+// CHECK-NOT: coalesce_hint
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+func.func @load_broadcast_offsets_no_hint(%ptr: i64) -> vector<32xf32> {
+  %offsets = arith.constant dense<0> : vector<32xindex>
+  %mask = arith.constant dense<true> : vector<32xi1>
+  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+  return %v : vector<32xf32>
+}
+
+// -----
+// Explicit chunk_size > 1: precondition fails, no hint stamped (regardless
+// of analysis result).
+// CHECK-LABEL: gpu.func @load_explicit_chunk_no_hint(
+// CHECK: xegpu.load
+// CHECK-NOT: coalesce_hint
+// CHECK-SAME: <{chunk_size = 2 : i64}>
+gpu.module @kernel_explicit_chunk [#xevm.target<chip = "pvc">] {
+  gpu.func @load_explicit_chunk_no_hint(%ptr: i64) -> vector<32x2xf32> {
+    %offsets = vector.step : vector<32xindex>
+    %mask = arith.constant dense<true> : vector<32xi1>
+    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 2 : i64}>
+        : i64, vector<32xindex>, vector<32xi1> -> vector<32x2xf32>
+    gpu.return %v : vector<32x2xf32>
+  }
+}
+
+// -----
+// Store with vector.step offsets: hint stamped with factor = 2 on the store
+// op.
+// CHECK-LABEL: func.func @store_step_hint(
+// CHECK: xegpu.store
+// CHECK-SAME: {xegpu.coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}
+// CHECK-SAME: : vector<32xf32>, i64, vector<32xindex>, vector<32xi1>
+// CHECK-NOT: lane_data
+func.func @store_step_hint(%ptr: i64, %v: vector<32xf32>) {
+  %offsets = vector.step : vector<32xindex>
+  %mask = arith.constant dense<true> : vector<32xi1>
+  xegpu.store %v, %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : vector<32xf32>, i64, vector<32xindex>, vector<32xi1>
+  return
+}
diff --git a/mlir/test/Dialect/XeGPU/coalesce-gather-scatter.mlir b/mlir/test/Dialect/XeGPU/coalesce-gather-scatter.mlir
new file mode 100644
index 0000000000000..171879f18ad87
--- /dev/null
+++ b/mlir/test/Dialect/XeGPU/coalesce-gather-scatter.mlir
@@ -0,0 +1,476 @@
+// RUN: mlir-opt -split-input-file -test-xegpu-coalesce-gather-scatter %s | FileCheck %s
+// RUN: mlir-opt -split-input-file -test-xegpu-coalesce-gather-scatter="max-chunk-size=4" %s | FileCheck --check-prefix=CHECK4 %s
+
+// -----
+// vector.step offsets -> stride 1, fully coalescible.
+// 32 lanes, subgroup_size = 16 (default) -> lane_layout = 16, perLane = 2,
+// max-chunk-size = 8 -> bound = min(32, 8, 2) = 2 -> lane_data = 2.
+// The trivial `chunk_size = 1` attribute is dropped on success since the
+// new lane_data FCD > 1 supersedes it.
+// CHECK-LABEL: func.func @load_step_offsets(
+// CHECK:   %[[STEP:.*]] = vector.step : vector<32xindex>
+// CHECK:   %[[LOAD:.*]] = xegpu.load
+// CHECK-NOT: chunk_size
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [32], lane_layout = [16], lane_data = [2]>
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+// CHECK:   return %[[LOAD]]
+func.func @load_step_offsets(%ptr: i64) -> vector<32xf32> {
+  %offsets = vector.step : vector<32xindex>
+  %mask = arith.constant dense<true> : vector<32xi1>
+  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+  return %v : vector<32xf32>
+}
+
+// -----
+// max-chunk-size = 4: same lane_layout / lane_data because the bound
+// `min(contiguity = 32, budget = 4, perLane = 2) = 2` already saturates at
+// perLane.
+// CHECK4-LABEL: func.func @load_step_offsets_chunk4(
+// CHECK4: xegpu.load
+// CHECK4-SAME: layout = #xegpu.layout<inst_data = [32], lane_layout = [16], lane_data = [2]>
+// CHECK4-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+func.func @load_step_offsets_chunk4(%ptr: i64) -> vector<32xf32> {
+  %offsets = vector.step : vector<32xindex>
+  %mask = arith.constant dense<true> : vector<32xi1>
+  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+  return %v : vector<32xf32>
+}
+
+// -----
+// Dense constant arithmetic progression with stride 1.
+// CHECK-LABEL: func.func @load_dense_ap_offsets(
+// CHECK: xegpu.load
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [32], lane_layout = [16], lane_data = [2]>
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xi32>
+func.func @load_dense_ap_offsets(%ptr: i64) -> vector<32xi32> {
+  %offsets = arith.constant dense<[
+    0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
+    16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
+  ]> : vector<32xindex>
+  %mask = arith.constant dense<true> : vector<32xi1>
+  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : i64, vector<32xindex>, vector<32xi1> -> vector<32xi32>
+  return %v : vector<32xi32>
+}
+
+// -----
+// Non-stride-1 (stride 4) offsets: not contiguous, no layout attached.
+// CHECK-LABEL: func.func @load_stride4_unchanged(
+// CHECK: xegpu.load
+// CHECK-NOT: lane_data
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+func.func @load_stride4_unchanged(%ptr: i64) -> vector<32xf32> {
+  %c4 = arith.constant 4 : index
+  %step = vector.step : vector<32xindex>
+  %splat = vector.broadcast %c4 : index to vector<32xindex>
+  %offsets = arith.muli %step, %splat : vector<32xindex>
+  %mask = arith.constant dense<true> : vector<32xi1>
+  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+  return %v : vector<32xf32>
+}
+
+// -----
+// Non-uniform mask: not coalesced.
+// CHECK-LABEL: func.func @load_partial_mask_unchanged(
+// CHECK: xegpu.load
+// CHECK-NOT: lane_data
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+func.func @load_partial_mask_unchanged(%ptr: i64, %mask: vector<32xi1>) -> vector<32xf32> {
+  %offsets = vector.step : vector<32xindex>
+  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+  return %v : vector<32xf32>
+}
+
+// -----
+// Store with vector.step offsets coalesces.
+// CHECK-LABEL: func.func @store_step_offsets(
+// CHECK: xegpu.store
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [32], lane_layout = [16], lane_data = [2]>
+// CHECK-SAME: : vector<32xf32>, i64, vector<32xindex>, vector<32xi1>
+func.func @store_step_offsets(%ptr: i64, %v: vector<32xf32>) {
+  %offsets = vector.step : vector<32xindex>
+  %mask = arith.constant dense<true> : vector<32xi1>
+  xegpu.store %v, %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : vector<32xf32>, i64, vector<32xindex>, vector<32xi1>
+  return
+}
+
+// -----
+// Store with all-equal offsets is left alone (ambiguous semantics).
+// CHECK-LABEL: func.func @store_broadcast_offsets_unchanged(
+// CHECK: xegpu.store
+// CHECK-NOT: lane_data
+// CHECK-SAME: vector<32xf32>, i64, vector<32xindex>, vector<32xi1>
+func.func @store_broadcast_offsets_unchanged(%ptr: i64, %v: vector<32xf32>) {
+  %offsets = arith.constant dense<0> : vector<32xindex>
+  %mask = arith.constant dense<true> : vector<32xi1>
+  xegpu.store %v, %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : vector<32xf32>, i64, vector<32xindex>, vector<32xi1>
+  return
+}
+
+// -----
+// memref-source variant of load coalesces too.
+// CHECK-LABEL: func.func @load_memref_step(
+// CHECK: xegpu.load
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [32], lane_layout = [16], lane_data = [2]>
+// CHECK-SAME: : memref<1024xf32>, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+func.func @load_memref_step(%m: memref<1024xf32>) -> vector<32xf32> {
+  %offsets = vector.step : vector<32xindex>
+  %mask = arith.constant dense<true> : vector<32xi1>
+  %v = xegpu.load %m[%offsets], %mask <{chunk_size = 1 : i64}>
+      : memref<1024xf32>, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+  return %v : vector<32xf32>
+}
+
+// -----
+// 2-D offsets with leading unit dim: inner dim treated as lane dim.
+// vector<1x32xindex> stride-1, subgroup_size = 16 ->
+// lane_layout = [1, 16], lane_data = [1, 2].
+// CHECK-LABEL: func.func @load_2d_leading_unit(
+// CHECK: xegpu.load
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [1, 32], lane_layout = [1, 16], lane_data = [1, 2]>
+// CHECK-SAME: : i64, vector<1x32xindex>, vector<1x32xi1> -> vector<1x32xf32>
+func.func @load_2d_leading_unit(%ptr: i64) -> vector<1x32xf32> {
+  %step = vector.step : vector<32xindex>
+  %offsets = vector.shape_cast %step : vector<32xindex> to vector<1x32xindex>
+  %mask = arith.constant dense<true> : vector<1x32xi1>
+  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : i64, vector<1x32xindex>, vector<1x32xi1> -> vector<1x32xf32>
+  return %v : vector<1x32xf32>
+}
+
+// -----
+// True 2-D dense AP: each row stride-1, inner = 16 = subgroup_size, so
+// perLane = 1 and there's no room for lane_data > 1. The pass leaves the
+// op alone; the wider variant `load_2d_dense_ap_wide` below exercises the
+// coalescing path.
+// CHECK-LABEL: func.func @load_2d_dense_ap(
+// CHECK: xegpu.load
+// CHECK-NOT: lane_data
+// CHECK-SAME: : i64, vector<2x16xindex>, vector<2x16xi1> -> vector<2x16xf32>
+func.func @load_2d_dense_ap(%ptr: i64) -> vector<2x16xf32> {
+  %offsets = arith.constant dense<[
+    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15],
+    [16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]
+  ]> : vector<2x16xindex>
+  %mask = arith.constant dense<true> : vector<2x16xi1>
+  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : i64, vector<2x16xindex>, vector<2x16xi1> -> vector<2x16xf32>
+  return %v : vector<2x16xf32>
+}
+
+// -----
+// True 2-D dense AP with inner > subgroup_size: each row stride-1 across 32
+// lanes. With subgroup_size = 16 (from #xevm.target chip = "pvc"),
+// lane_layout[inner] = 16, perLane = 32 / 16 = 2, contiguity[inner] = 32,
+// budget = max-chunk-size / 1 = 8. bound = min(32, 8, 2) = 2 -> lane_data = 2.
+// CHECK-LABEL: gpu.func @load_2d_dense_ap_wide(
+// CHECK: xegpu.load
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [1, 32], lane_layout = [1, 16], lane_data = [1, 2]>
+// CHECK-SAME: : i64, vector<2x32xindex>, vector<2x32xi1> -> vector<2x32xf32>
+gpu.module @kernel [#xevm.target<chip = "pvc">] {
+  gpu.func @load_2d_dense_ap_wide(%ptr: i64) -> vector<2x32xf32> {
+    %offsets = arith.constant dense<[
+      [ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15,
+       16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31],
+      [32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
+       48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63]
+    ]> : vector<2x32xindex>
+    %mask = arith.constant dense<true> : vector<2x32xi1>
+    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+        : i64, vector<2x32xindex>, vector<2x32xi1> -> vector<2x32xf32>
+    gpu.return %v : vector<2x32xf32>
+  }
+}
+
+// -----
+// 1-D step reshape_cast'ed to 2x32: inner extent 32 > subgroup_size = 16, so
+// the lane_layout-first rule picks lane_layout[inner] = 16 (perLane = 2),
+// then takes lane_data[inner] = min(contiguity = 32, budget = 8, perLane = 2)
+// rounded down to a power-of-2 divisor of perLane => 2.
+// CHECK-LABEL: gpu.func @load_2x32_step_shape_cast(
+// CHECK: xegpu.load
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [1, 32], lane_layout = [1, 16], lane_data = [1, 2]>
+// CHECK-SAME: : i64, vector<2x32xindex>, vector<2x32xi1> -> vector<2x32xf32>
+gpu.module @kernel_2x32 [#xevm.target<chip = "pvc">] {
+  gpu.func @load_2x32_step_shape_cast(%ptr: i64) -> vector<2x32xf32> {
+    %step = vector.step : vector<64xindex>
+    %offsets = vector.shape_cast %step : vector<64xindex> to vector<2x32xindex>
+    %mask = arith.constant dense<true> : vector<2x32xi1>
+    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+        : i64, vector<2x32xindex>, vector<2x32xi1> -> vector<2x32xf32>
+    gpu.return %v : vector<2x32xf32>
+  }
+}
+
+// -----
+// Negative: 2-D dense values where inner row is not stride-1 AP. Should
+// not coalesce.
+// CHECK-LABEL: func.func @load_2d_non_ap_unchanged(
+// CHECK: xegpu.load
+// CHECK-NOT: lane_data
+// CHECK-SAME: : i64, vector<2x16xindex>, vector<2x16xi1> -> vector<2x16xf32>
+func.func @load_2d_non_ap_unchanged(%ptr: i64) -> vector<2x16xf32> {
+  %offsets = arith.constant dense<[
+    [0, 1, 2, 3, 4, 5, 6, 7, 100, 9, 10, 11, 12, 13, 14, 15],
+    [16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]
+  ]> : vector<2x16xindex>
+  %mask = arith.constant dense<true> : vector<2x16xi1>
+  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : i64, vector<2x16xindex>, vector<2x16xi1> -> vector<2x16xf32>
+  return %v : vector<2x16xf32>
+}
+
+// -----
+// "Reduction kernel" pattern with inner > subgroup_size: 2-D offsets built
+// from `transpose(broadcast(rowOffsets))` + `broadcast(step)`. The transpose
+// supplies inner-dim constancy, the broadcast(step) supplies inner-dim
+// contiguity, and the addi recovers contiguity through visitAddSub. With
+// inner = 32 and subgroup_size = 16 this picks lane_layout = [1, 16],
+// lane_data = [1, 2].
+// CHECK-LABEL: gpu.func @load_reduction_pattern_wide(
+// CHECK: xegpu.load
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [1, 32], lane_layout = [1, 16], lane_data = [1, 2]>
+// CHECK-SAME: : i64, vector<2x32xindex>, vector<2x32xi1> -> vector<2x32xf32>
+gpu.module @kernel_reduction [#xevm.target<chip = "pvc">] {
+  gpu.func @load_reduction_pattern_wide(%ptr: i64, %row0: index, %row1: index)
+      -> vector<2x32xf32> {
+    // Per-row base offsets.
+    %r0   = vector.broadcast %row0 : index to vector<1xindex>
+    %r1   = vector.broadcast %row1 : index to vector<1xindex>
+    %rows = vector.shuffle %r0, %r1 [0, 1] : vector<1xindex>, vector<1xindex>
+    // Inner stride-1 step.
+    %step = vector.step : vector<32xindex>
+    // Build 2x32 offsets: broadcast rows -> transpose -> add broadcast(step).
+    %rowsBc = vector.broadcast %rows
+            : vector<2xindex> to vector<32x2xindex>
+    %rowsT  = vector.transpose %rowsBc, [1, 0]
+            : vector<32x2xindex> to vector<2x32xindex>
+    %cols2  = vector.broadcast %step
+            : vector<32xindex> to vector<2x32xindex>
+    %off    = arith.addi %rowsT, %cols2 : vector<2x32xindex>
+    %mask   = arith.constant dense<true> : vector<2x32xi1>
+    %v = xegpu.load %ptr[%off], %mask <{chunk_size = 1 : i64}>
+        : i64, vector<2x32xindex>, vector<2x32xi1> -> vector<2x32xf32>
+    gpu.return %v : vector<2x32xf32>
+  }
+}
+
+// -----
+// `divui` by a uniform constant equal to the inner stride recovers stride-1
+// contiguity. Here offsets = [0,2,4,…,62] / 2 = [0,1,…,31], 32 lanes
+// against subgroup_size = 16 -> lane_layout = 16, perLane = 2, contiguity
+// = 32, bound = min(32, 8, 2) = 2 -> lane_data = 2.
+// CHECK-LABEL: gpu.func @load_divui_recovers_contiguity(
+// CHECK: xegpu.load
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [32], lane_layout = [16], lane_data = [2]>
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+gpu.module @kernel_divui [#xevm.target<chip = "pvc">] {
+  gpu.func @load_divui_recovers_contiguity(%ptr: i64) -> vector<32xf32> {
+    %even = arith.constant dense<[
+       0,  2,  4,  6,  8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30,
+      32, 34, 36, 38, 40, 42, 44, 46, 48, 50, 52, 54, 56, 58, 60, 62]>
+        : vector<32xindex>
+    %c2   = arith.constant dense<2> : vector<32xindex>
+    %offsets = arith.divui %even, %c2 : vector<32xindex>
+    %mask = arith.constant dense<true> : vector<32xi1>
+    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+        : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+    gpu.return %v : vector<32xf32>
+  }
+}
+
+// -----
+// `divsi` by a uniform constant equal to the inner stride: same recovery as
+// the divui case.
+// CHECK-LABEL: gpu.func @load_divsi_recovers_contiguity(
+// CHECK: xegpu.load
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [32], lane_layout = [16], lane_data = [2]>
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+gpu.module @kernel_divsi [#xevm.target<chip = "pvc">] {
+  gpu.func @load_divsi_recovers_contiguity(%ptr: i64) -> vector<32xf32> {
+    %even = arith.constant dense<[
+       0,  2,  4,  6,  8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30,
+      32, 34, 36, 38, 40, 42, 44, 46, 48, 50, 52, 54, 56, 58, 60, 62]>
+        : vector<32xindex>
+    %c2   = arith.constant dense<2> : vector<32xindex>
+    %offsets = arith.divsi %even, %c2 : vector<32xindex>
+    %mask = arith.constant dense<true> : vector<32xi1>
+    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+        : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+    gpu.return %v : vector<32xf32>
+  }
+}
+
+// -----
+// `remui` by a constant that divides the inner stride: every element of a
+// row is the same residue class -> inner-dim constant. The decision picks
+// the broadcast case, which is currently disabled, so the load is left
+// alone (no layout, no chunk_size change).
+// CHECK-LABEL: gpu.func @load_remui_inner_uniform(
+// CHECK: xegpu.load
+// CHECK-NOT: lane_data
+// CHECK-SAME: : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
+gpu.module @kernel_remui [#xevm.target<chip = "pvc">] {
+  gpu.func @load_remui_inner_uniform(%ptr: i64) -> vector<16xf32> {
+    %even = arith.constant dense<[
+      0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30]>
+        : vector<16xindex>
+    %c2   = arith.constant dense<2> : vector<16xindex>
+    // (even % 2) is uniformly 0 along the inner dim.
+    %offsets = arith.remui %even, %c2 : vector<16xindex>
+    %mask = arith.constant dense<true> : vector<16xi1>
+    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+        : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
+    gpu.return %v : vector<16xf32>
+  }
+}
+
+// -----
+// `andi` with a power-of-two-minus-one mask is `% (1 << k)`. With stride 2
+// and mask 1 (= 2-1), every result element is uniform: same broadcast case
+// as remui above, currently disabled, so the load is left alone.
+// CHECK-LABEL: gpu.func @load_andi_inner_uniform(
+// CHECK: xegpu.load
+// CHECK-NOT: lane_data
+// CHECK-SAME: : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
+gpu.module @kernel_andi [#xevm.target<chip = "pvc">] {
+  gpu.func @load_andi_inner_uniform(%ptr: i64) -> vector<16xf32> {
+    %even = arith.constant dense<[
+      0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30]>
+        : vector<16xindex>
+    %m = arith.constant dense<1> : vector<16xindex>
+    %offsets = arith.andi %even, %m : vector<16xindex>
+    %mask = arith.constant dense<true> : vector<16xi1>
+    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+        : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
+    gpu.return %v : vector<16xf32>
+  }
+}
+
+// -----
+// `shli` by a constant scales the inner stride. step << 1 -> stride 2,
+// which on its own is not coalescible (no divui follows), so we get no
+// layout attached even when there's room (vector<32>, perLane=2).
+// CHECK-LABEL: gpu.func @load_shli_unchanged(
+// CHECK: xegpu.load
+// CHECK-NOT: lane_data
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+gpu.module @kernel_shli [#xevm.target<chip = "pvc">] {
+  gpu.func @load_shli_unchanged(%ptr: i64) -> vector<32xf32> {
+    %step = vector.step : vector<32xindex>
+    %k = arith.constant dense<1> : vector<32xindex>
+    %offsets = arith.shli %step, %k : vector<32xindex>
+    %mask = arith.constant dense<true> : vector<32xi1>
+    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+        : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+    gpu.return %v : vector<32xf32>
+  }
+}
+
+// -----
+// `shli` followed by `shrui` cancels: `(step << 1) >> 1` has innerStride
+// scaled to 2 then divided by 2 -> stride 1 again. With inner = 32 and
+// subgroup_size = 16 we coalesce by lane_data = 2.
+// CHECK-LABEL: gpu.func @load_shli_then_shrui_recovers(
+// CHECK: xegpu.load
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [32], lane_layout = [16], lane_data = [2]>
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+gpu.module @kernel_shli_shrui [#xevm.target<chip = "pvc">] {
+  gpu.func @load_shli_then_shrui_recovers(%ptr: i64) -> vector<32xf32> {
+    %step = vector.step : vector<32xindex>
+    %k = arith.constant dense<1> : vector<32xindex>
+    %doubled = arith.shli %step, %k : vector<32xindex>
+    %offsets = arith.shrui %doubled, %k : vector<32xindex>
+    %mask = arith.constant dense<true> : vector<32xi1>
+    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+        : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+    gpu.return %v : vector<32xf32>
+  }
+}
+
+// -----
+// `arith.select` between two AP arms with the same inner-dim properties:
+// the result inherits the meet of the two arms. Both arms here are
+// stride-1 step + a per-arm constant base, so the select preserves
+// inner-dim contiguity = 32 -> coalesces with lane_data = 2.
+// CHECK-LABEL: gpu.func @load_select_two_aps(
+// CHECK: xegpu.load
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [32], lane_layout = [16], lane_data = [2]>
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+gpu.module @kernel_select [#xevm.target<chip = "pvc">] {
+  gpu.func @load_select_two_aps(%ptr: i64, %cond: i1) -> vector<32xf32> {
+    %step = vector.step : vector<32xindex>
+    %a = arith.constant dense<0>  : vector<32xindex>
+    %b = arith.constant dense<64> : vector<32xindex>
+    %baseA = arith.addi %step, %a : vector<32xindex>
+    %baseB = arith.addi %step, %b : vector<32xindex>
+    %offsets = arith.select %cond, %baseA, %baseB : vector<32xindex>
+    %mask = arith.constant dense<true> : vector<32xi1>
+    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+        : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+    gpu.return %v : vector<32xf32>
+  }
+}
+
+// -----
+// `divui` by a constant that does NOT divide the inner stride: stride 2 / 3
+// is not exact, so we conservatively give up. No layout attached.
+// CHECK-LABEL: gpu.func @load_divui_non_divisor_unchanged(
+// CHECK: xegpu.load
+// CHECK-NOT: lane_data
+// CHECK-SAME: : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
+gpu.module @kernel_divui_neg [#xevm.target<chip = "pvc">] {
+  gpu.func @load_divui_non_divisor_unchanged(%ptr: i64) -> vector<16xf32> {
+    %even = arith.constant dense<[
+      0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30]>
+        : vector<16xindex>
+    %c3   = arith.constant dense<3> : vector<16xindex>
+    %offsets = arith.divui %even, %c3 : vector<16xindex>
+    %mask = arith.constant dense<true> : vector<16xi1>
+    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+        : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
+    gpu.return %v : vector<16xf32>
+  }
+}
+
+// -----
+// 2-D store with step + shape_cast offsets coalesces too.
+// CHECK-LABEL: func.func @store_2d_step(
+// CHECK: xegpu.store
+// CHECK-SAME: layout = #xegpu.layout<inst_data = [1, 32], lane_layout = [1, 16], lane_data = [1, 2]>
+// CHECK-SAME: : vector<1x32xf32>, i64, vector<1x32xindex>, vector<1x32xi1>
+func.func @store_2d_step(%ptr: i64, %v: vector<1x32xf32>) {
+  %step = vector.step : vector<32xindex>
+  %offsets = vector.shape_cast %step : vector<32xindex> to vector<1x32xindex>
+  %mask = arith.constant dense<true> : vector<1x32xi1>
+  xegpu.store %v, %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+      : vector<1x32xf32>, i64, vector<1x32xindex>, vector<1x32xi1>
+  return
+}
+
+// -----
+// An op that already declares chunk_size = 2 is left alone: the verifier
+// requires a particular value/mask shape relationship for chunked ops, so
+// the pass conservatively skips when an explicit chunk_size > 1 is set
+// (a downstream pass has already committed to that per-lane chunked
+// access).
+// CHECK-LABEL: gpu.func @load_explicit_chunk_unchanged(
+// CHECK: xegpu.load
+// CHECK-NOT: lane_data
+// CHECK-SAME: <{chunk_size = 2 : i64}>
+// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32x2xf32>
+gpu.module @kernel_explicit_chunk [#xevm.target<chip = "pvc">] {
+  gpu.func @load_explicit_chunk_unchanged(%ptr: i64) -> vector<32x2xf32> {
+    %offsets = vector.step : vector<32xindex>
+    %mask = arith.constant dense<true> : vector<32xi1>
+    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 2 : i64}>
+        : i64, vector<32xindex>, vector<32xi1> -> vector<32x2xf32>
+    gpu.return %v : vector<32x2xf32>
+  }
+}
diff --git a/mlir/test/lib/Dialect/XeGPU/TestXeGPUTransforms.cpp b/mlir/test/lib/Dialect/XeGPU/TestXeGPUTransforms.cpp
index 5c3721630837d..b71bd6a0be618 100644
--- a/mlir/test/lib/Dialect/XeGPU/TestXeGPUTransforms.cpp
+++ b/mlir/test/lib/Dialect/XeGPU/TestXeGPUTransforms.cpp
@@ -446,6 +446,48 @@ struct TestXeGPULayoutInterface
   }
 };
 
+struct TestXeGPUCoalesceGatherScatter
+    : public PassWrapper<TestXeGPUCoalesceGatherScatter, OperationPass<>> {
+  MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(TestXeGPUCoalesceGatherScatter)
+
+  StringRef getArgument() const final {
+    return "test-xegpu-coalesce-gather-scatter";
+  }
+
+  StringRef getDescription() const final {
+    return "Test the XeGPU coalesce-gather-scatter analysis + apply APIs.";
+  }
+
+  void getDependentDialects(::mlir::DialectRegistry &registry) const override {
+    registry.insert<arith::ArithDialect>();
+    registry.insert<vector::VectorDialect>();
+    registry.insert<xegpu::XeGPUDialect>();
+  }
+
+  TestXeGPUCoalesceGatherScatter() = default;
+  TestXeGPUCoalesceGatherScatter(const TestXeGPUCoalesceGatherScatter &pass)
+      : PassWrapper(pass) {}
+
+  Option<unsigned> maxChunkSize{
+      *this, "max-chunk-size",
+      llvm::cl::desc("Upper bound on the produced lane_data FCD."),
+      llvm::cl::init(8)};
+
+  Option<bool> analyzeOnly{
+      *this, "analyze-only",
+      llvm::cl::desc("Only run the analysis (stamp xegpu.coalesce_hint "
+                     "attributes); do not apply."),
+      llvm::cl::init(false)};
+
+  void runOnOperation() override {
+    xegpu::CoalesceGatherScatterAnalysisOptions options;
+    options.maxChunkSize = maxChunkSize;
+    xegpu::runCoalesceGatherScatterAnalysis(getOperation(), options);
+    if (!analyzeOnly)
+      xegpu::applyCoalesceGatherScatterHints(getOperation());
+  }
+};
+
 } // namespace
 
 namespace mlir {
@@ -458,6 +500,7 @@ void registerTestXeGPULowerings() {
   PassRegistration<TestXeGPUPropagateLayouts>();
   PassRegistration<TestXeGPUResolveLayoutConflicts>();
   PassRegistration<TestXeGPUArrayLengthOptimization>();
+  PassRegistration<TestXeGPUCoalesceGatherScatter>();
 }
 } // namespace test
 } // namespace mlir

>From 5ef04fb11afa1215b7f1da6f99099863da88e764 Mon Sep 17 00:00:00 2001
From: "Shahneous Bari, Md Abdullah" <md.abdullah.shahneous.bari at intel.com>
Date: Wed, 10 Jun 2026 20:05:21 +0000
Subject: [PATCH 2/7] Address review comments.

---
 .../mlir/Dialect/XeGPU/IR/XeGPUAttrs.td       | 26 +++--
 .../Dialect/XeGPU/Transforms/Transforms.h     | 29 +++---
 .../Transforms/XeGPUCoalesceGatherScatter.cpp | 99 ++++++++++---------
 .../coalesce-gather-scatter-analyze.mlir      | 33 ++-----
 .../XeGPU/coalesce-gather-scatter.mlir        | 74 +++++---------
 .../lib/Dialect/XeGPU/TestXeGPUTransforms.cpp |  2 +-
 6 files changed, 107 insertions(+), 156 deletions(-)

diff --git a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
index 50b67cefafa31..292302a74934c 100644
--- a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
+++ b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
@@ -933,24 +933,22 @@ def XeGPU_CoalesceHintAttr : XeGPUAttr<"CoalesceHint", "coalesce_hint"> {
   let description = [{
     `CoalesceHintAttr` is a discardable attribute attached by the
     coalesce-gather-scatter analysis to `xegpu.load` / `xegpu.store` ops
-    whose offsets describe a contiguous-per-lane access pattern. It records
-    the chosen `lane_data` factor along the fastest-changing dim (FCD) of
-    the value vector. The `lane_layout` along the FCD is *not* stored here;
-    it is re-derived at apply time from the op's offsets inner extent and
-    the chip's subgroup size, mirroring `xegpu-propagate-layout`'s default
-    rule.
-
-    The attribute is consumed and removed by `applyCoalesceGatherScatterHint`,
-    which installs an equivalent `xegpu.layout` on the op. A producer that
-    decides not to apply the hint (for example, a propagator that detects a
-    conflict with an anchor-driven layout) should remove the attribute
-    rather than leave it dangling.
+    whose offsets describe a coalescible access pattern. `factor` is the
+    number of contiguous elements grouped together along the innermost
+    (fastest-changing) dimension of the access.
+
+    The attribute is purely advisory: it records the analysis's decision
+    without committing to a layout. A consumer turns `factor` into a concrete
+    `xegpu.layout` (deriving the `lane_layout` / `lane_data` split from the
+    op's offsets inner extent and the chip's subgroup size) and removes the
+    hint. A consumer that decides not to act on the hint (for example, a
+    propagator that detects a conflict with an anchor-driven layout) should
+    likewise remove the attribute rather than leave it dangling.
 
     Example:
     ```mlir
     %v = xegpu.load %ptr[%offsets], %mask
-            <{chunk_size = 1 : i64,
-              xegpu.coalesce_hint = #xegpu.coalesce_hint<factor = 4>}>
+            {xegpu.coalesce_hint = #xegpu.coalesce_hint<factor = 4>}
             : i64, vector<64xindex>, vector<64xi1> -> vector<64xf32>
     ```
   }];
diff --git a/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h b/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h
index 6046f78dd59c0..fb560acc3dc48 100644
--- a/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h
+++ b/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h
@@ -91,33 +91,26 @@ inline StringRef getCoalesceHintAttrName() { return "xegpu.coalesce_hint"; }
 
 /// Options controlling `runCoalesceGatherScatterAnalysis`.
 struct CoalesceGatherScatterAnalysisOptions {
-  /// Upper bound on the per-lane chunk size produced by coalescing. Mirrors
-  /// the `max-chunk-size` option of the original pass.
+  /// Upper bound on the number of contiguous elements grouped per lane by
+  /// coalescing. Mirrors the `max-chunk-size` option of the original pass.
   unsigned maxChunkSize = 8;
 };
 
 /// Run the AxisInfo-based coalescing analysis over `root` and stamp a
 /// `xegpu.coalesce_hint` attribute on every `xegpu.load` / `xegpu.store`
 /// the analysis classifies as coalescible. Ops with an existing non-empty
-/// `lane_data`, an explicit `chunk_size > 1`, or a non-uniform mask are
-/// skipped (no hint stamped).
-///
-/// This function performs no rewrite of its own; the hint is consumed by
-/// `applyCoalesceGatherScatterHint` (or a downstream pass that wants
-/// stronger control over when to honor the hint).
+/// `lane_data` or a non-uniform mask are skipped (no hint stamped). The hint
+/// is consumed downstream — either by `coalesceGatherScatter` or by a
+/// layout-propagation pass that reads the hint directly.
 void runCoalesceGatherScatterAnalysis(
     Operation *root, const CoalesceGatherScatterAnalysisOptions &options = {});
 
-/// Apply a stamped `xegpu.coalesce_hint` on `op`: install an equivalent
-/// `lane_layout` / `lane_data` / `inst_data` layout, drop a trivial
-/// `chunk_size = 1` attribute, and remove the hint. Idempotent — no-op if
-/// the op carries no hint. Returns `failure()` when the hint is malformed
-/// (e.g. attached to an op that isn't a gather/scatter).
-LogicalResult applyCoalesceGatherScatterHint(Operation *op);
-
-/// Walk `root` and apply coalesce hints on every op that carries one.
-/// Hints stamped on unsupported ops are silently dropped.
-void applyCoalesceGatherScatterHints(Operation *root);
+/// Walk `root` and turn every stamped `xegpu.coalesce_hint` into an
+/// equivalent `lane_layout` / `lane_data` / `inst_data` layout on its op,
+/// then remove the hint. A hint on an op that cannot be coalesced (e.g. one
+/// that isn't a gather/scatter) is simply dropped. Pairs with
+/// `runCoalesceGatherScatterAnalysis`.
+void coalesceGatherScatter(Operation *root);
 
 /// Walk `root` and remove any leftover `xegpu.coalesce_hint` attributes —
 /// useful as a cleanup after a propagator has decided whether to honor each
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
index 669794f5c41fd..52f1b321be900 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
@@ -6,14 +6,18 @@
 //
 //===----------------------------------------------------------------------===//
 //
-// This pass coalesces neighbouring lanes of `xegpu.load` / `xegpu.store` ops
-// so that each lane handles `N` contiguous elements along the innermost
-// dimension. The decision is driven by a small XeGPU-local axis-info
-// dataflow analysis modeled on Triton's `AxisInfo` (`contiguity`,
-// `constancy`, `divisibility`) and is applied by attaching a
-// `lane_data` layout to the original op. The actual memory-message rewrite
-// is left to the downstream WG-to-SG / SG-to-Lane distribution passes,
-// which interpret `lane_data`.
+// This file implements the gather/scatter coalescing analysis. It decides
+// whether an `xegpu.load` / `xegpu.store` gather/scatter accesses `N`
+// contiguous elements per lane along the innermost dimension, and, if so,
+// stamps a `xegpu.coalesce_hint<factor = N>` attribute recording the chosen
+// factor. The decision is driven by a small XeGPU-local axis-info dataflow
+// analysis tracking per-axis `contiguity`, `constancy`, and `divisibility`.
+//
+// The analysis performs no rewrite. The hint it stamps is turned into a
+// `lane_data` layout by `coalesceGatherScatter` (used by the test pass) or
+// read directly by layout propagation; the actual memory-message rewrite is
+// left to the downstream WG-to-SG / SG-to-Lane distribution passes, which
+// interpret `lane_data`.
 //
 // The analysis tracks per-axis information for vectors of integer / index
 // type at any rank. The coalescing decision is computed against the
@@ -80,10 +84,15 @@ static constexpr int64_t kAxisInfoTop = 1LL << 30;
 ///   - `knownConstant`: scalar value if the entire vector is uniformly
 ///     known to be a single constant.
 ///   - `innerStride`: when set, every "row" along the innermost dimension
-///     is an arithmetic progression with this stride. Per-row base may
-///     differ across outer indices; per-row alignment is captured by
-///     `divisibility[innerDim]`. `innerStride = 1` implies stride-1
-///     contiguity; `innerStride = 0` implies inner-dim constancy.
+///     is an arithmetic progression with this (constant) step. This is the
+///     general form of which `contiguity`/`constancy` are the two special
+///     cases: `innerStride = 1` is the stride-1 contiguous case (and implies
+///     `contiguity[innerDim] > 1`), `innerStride = 0` is the all-equal case
+///     (implies `constancy[innerDim] > 1`), and any other value (e.g. 4) is a
+///     strided progression that neither `contiguity` nor `constancy` can
+///     represent — both read 1 there. Per-row base may differ across outer
+///     indices; per-row alignment is captured by `divisibility[innerDim]`.
+///     The coalescing decision only acts on the `innerStride = 1` case.
 ///
 /// Pessimistic / entry value: contiguity=1, constancy=1, divisibility=1,
 /// innerStride absent.
@@ -888,11 +897,10 @@ static int64_t largestPow2Divisor(int64_t numLanes, int64_t bound) {
 /// `XeGPUPropagateLayout`: `lane_layout[inner] = min(subgroupSize, inner)`,
 /// rounded down to a power-of-2 divisor of `inner`. The remaining lane
 /// budget then becomes `lane_data[inner] = inner / lane_layout`, capped by
-/// the per-lane chunk-size budget and the offsets contiguity.
+/// `maxChunkSize` and the offsets contiguity.
 static CoalesceDecision decide(const AxisInfo &info,
                                ArrayRef<int64_t> offsetsShape,
-                               int64_t origChunk, unsigned maxChunkSize,
-                               unsigned subgroupSize) {
+                               unsigned maxChunkSize, unsigned subgroupSize) {
   CoalesceDecision d;
   if (!info.isInitialized() || offsetsShape.empty())
     return d;
@@ -918,9 +926,7 @@ static CoalesceDecision decide(const AxisInfo &info,
   if (perLane < 2)
     return d; // already one element per lane, nothing to coalesce.
 
-  if (origChunk < 1)
-    origChunk = 1;
-  int64_t budget = static_cast<int64_t>(maxChunkSize) / origChunk;
+  int64_t budget = static_cast<int64_t>(maxChunkSize);
   if (budget < 2)
     return d;
 
@@ -969,11 +975,9 @@ static xegpu::LayoutAttr buildLaneDataLayout(MLIRContext *ctx, unsigned rank,
 }
 
 //===----------------------------------------------------------------------===//
-// Rewrites.
+// Analysis driver + hint apply.
 //===----------------------------------------------------------------------===//
 
-namespace {
-
 /// Look up the subgroup size from the enclosing gpu.module's xevm.target.
 /// Falls back to 16 when no target chip is found or the chip is unknown,
 /// matching the typical Intel Xe2 default. This keeps the pass usable on
@@ -986,7 +990,7 @@ static unsigned lookupSubgroupSize(Operation *op) {
 }
 
 /// Common analysis preconditions: vector offsets/value, all-true mask,
-/// no existing non-trivial lane_data, no explicit chunk_size > 1.
+/// no existing non-trivial lane_data.
 template <typename OpTy>
 static bool isCandidateForCoalesce(OpTy op) {
   auto offsetsTy = dyn_cast<VectorType>(op.getOffsets().getType());
@@ -999,8 +1003,6 @@ static bool isCandidateForCoalesce(OpTy op) {
   if (auto layout = op.getLayoutAttr())
     if (!layout.getEffectiveLaneDataAsInt().empty())
       return false;
-  if (op.getChunkSizeAttr() && op.getChunkSize().value_or(1) > 1)
-    return false;
   return true;
 }
 
@@ -1075,9 +1077,8 @@ static void analyzeAndStampHint(OpTy op, DataFlowSolver &solver,
   const auto *lat = solver.lookupState<AxisInfoLattice>(op.getOffsets());
   if (!lat || !lat->getValue().isInitialized())
     return;
-  int64_t origChunk = static_cast<int64_t>(op.getChunkSize().value_or(1));
-  auto d = decide(lat->getValue(), offsetsTy.getShape(), origChunk,
-                  maxChunkSize, subgroupSize);
+  auto d = decide(lat->getValue(), offsetsTy.getShape(), maxChunkSize,
+                  subgroupSize);
   if (d.kind != CoalesceDecision::Kind::Chunked)
     return;
   auto hint = xegpu::CoalesceHintAttr::get(op.getContext(), d.factor);
@@ -1086,9 +1087,9 @@ static void analyzeAndStampHint(OpTy op, DataFlowSolver &solver,
 
 /// Apply a stamped hint on `op`: build a lane_layout/lane_data/inst_data
 /// layout from the hint's `factor` and the op's offsets inner extent +
-/// chip-derived subgroup size, install it, drop a trivial `chunk_size = 1`,
-/// and remove the hint. Returns success on apply (or no-op when no hint),
-/// failure when the hint is malformed.
+/// chip-derived subgroup size, install it, and remove the hint. Returns
+/// success on apply (or no-op when no hint), failure when the hint is
+/// malformed.
 template <typename OpTy>
 static LogicalResult applyHintOnOp(OpTy op) {
   auto hint = op->template getAttrOfType<xegpu::CoalesceHintAttr>(
@@ -1111,16 +1112,25 @@ static LogicalResult applyHintOnOp(OpTy op) {
 
   auto layout = buildLaneDataLayout(op.getContext(), valueTy.getRank(),
                                     laneLayout, factor);
-  int64_t origChunk = static_cast<int64_t>(op.getChunkSize().value_or(1));
-  bool dropChunk = op.getChunkSizeAttr() && origChunk == 1 && factor > 1;
   op.setLayoutAttr(layout);
-  if (dropChunk)
-    op.removeChunkSizeAttr();
   op->removeAttr(xegpu::getCoalesceHintAttrName());
   return success();
 }
 
-} // namespace
+/// Apply a stamped `xegpu.coalesce_hint` on a single op. A hint on an op
+/// that cannot be coalesced (anything other than a gather/scatter) is just
+/// dropped. Returns failure only when the hint is malformed for a real
+/// gather/scatter.
+static LogicalResult applyHintOnOp(Operation *op) {
+  if (auto load = dyn_cast<xegpu::LoadGatherOp>(op))
+    return applyHintOnOp(load);
+  if (auto store = dyn_cast<xegpu::StoreScatterOp>(op))
+    return applyHintOnOp(store);
+  // Hint attached to an unsupported op: silently drop it.
+  if (op->hasAttr(xegpu::getCoalesceHintAttrName()))
+    op->removeAttr(xegpu::getCoalesceHintAttrName());
+  return success();
+}
 
 } // namespace
 
@@ -1136,6 +1146,10 @@ void mlir::xegpu::runCoalesceGatherScatterAnalysis(
   if (failed(solver.initializeAndRun(root)))
     return;
 
+  // The solver computed AxisInfo for the whole region in the single
+  // `initializeAndRun` above; offsets shared by several gather/scatter ops are
+  // analyzed only once. This walk is just per-op point lookups into that
+  // result (no re-analysis), turning each cached fact into a hint.
   root->walk([&](Operation *op) {
     if (auto load = dyn_cast<xegpu::LoadGatherOp>(op))
       analyzeAndStampHint(load, solver, options.maxChunkSize);
@@ -1144,21 +1158,10 @@ void mlir::xegpu::runCoalesceGatherScatterAnalysis(
   });
 }
 
-LogicalResult mlir::xegpu::applyCoalesceGatherScatterHint(Operation *op) {
-  if (auto load = dyn_cast<xegpu::LoadGatherOp>(op))
-    return applyHintOnOp(load);
-  if (auto store = dyn_cast<xegpu::StoreScatterOp>(op))
-    return applyHintOnOp(store);
-  // Hint attached to an unsupported op: silently drop it.
-  if (op->hasAttr(xegpu::getCoalesceHintAttrName()))
-    op->removeAttr(xegpu::getCoalesceHintAttrName());
-  return success();
-}
-
-void mlir::xegpu::applyCoalesceGatherScatterHints(Operation *root) {
+void mlir::xegpu::coalesceGatherScatter(Operation *root) {
   root->walk([&](Operation *op) {
     if (op->hasAttr(xegpu::getCoalesceHintAttrName()))
-      (void)applyCoalesceGatherScatterHint(op);
+      (void)applyHintOnOp(op);
   });
 }
 
diff --git a/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir b/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir
index ddfcf6e506c94..4b528704607a9 100644
--- a/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir
+++ b/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir
@@ -2,22 +2,20 @@
 // RUN:   -test-xegpu-coalesce-gather-scatter="analyze-only=true" %s | FileCheck %s
 
 // Analyze-only mode: stamps `xegpu.coalesce_hint` on coalescible ops and
-// leaves the layout / chunk_size unchanged. This test pins the hint
-// attribute contract that the apply API (and downstream propagator
-// integrations) consume.
+// leaves the layout unchanged. This test pins the hint attribute contract
+// that the apply API (and downstream propagator integrations) consume.
 
 // -----
 // 1-D vector.step, fully coalescible -> hint with factor = 2 stamped.
 // CHECK-LABEL: func.func @load_step_offsets(
 // CHECK: xegpu.load
-// CHECK-SAME: <{chunk_size = 1 : i64}>
 // CHECK-SAME: {xegpu.coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}
 // CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
 // CHECK-NOT: lane_data
 func.func @load_step_offsets(%ptr: i64) -> vector<32xf32> {
   %offsets = vector.step : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
-  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %ptr[%offsets], %mask
       : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
   return %v : vector<32xf32>
 }
@@ -33,7 +31,7 @@ func.func @load_2d_leading_unit(%ptr: i64) -> vector<1x32xf32> {
   %step = vector.step : vector<32xindex>
   %offsets = vector.shape_cast %step : vector<32xindex> to vector<1x32xindex>
   %mask = arith.constant dense<true> : vector<1x32xi1>
-  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %ptr[%offsets], %mask
       : i64, vector<1x32xindex>, vector<1x32xi1> -> vector<1x32xf32>
   return %v : vector<1x32xf32>
 }
@@ -50,7 +48,7 @@ func.func @load_stride4_no_hint(%ptr: i64) -> vector<32xf32> {
   %splat = vector.broadcast %c4 : index to vector<32xindex>
   %offsets = arith.muli %step, %splat : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
-  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %ptr[%offsets], %mask
       : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
   return %v : vector<32xf32>
 }
@@ -66,28 +64,11 @@ func.func @load_stride4_no_hint(%ptr: i64) -> vector<32xf32> {
 func.func @load_broadcast_offsets_no_hint(%ptr: i64) -> vector<32xf32> {
   %offsets = arith.constant dense<0> : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
-  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %ptr[%offsets], %mask
       : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
   return %v : vector<32xf32>
 }
 
-// -----
-// Explicit chunk_size > 1: precondition fails, no hint stamped (regardless
-// of analysis result).
-// CHECK-LABEL: gpu.func @load_explicit_chunk_no_hint(
-// CHECK: xegpu.load
-// CHECK-NOT: coalesce_hint
-// CHECK-SAME: <{chunk_size = 2 : i64}>
-gpu.module @kernel_explicit_chunk [#xevm.target<chip = "pvc">] {
-  gpu.func @load_explicit_chunk_no_hint(%ptr: i64) -> vector<32x2xf32> {
-    %offsets = vector.step : vector<32xindex>
-    %mask = arith.constant dense<true> : vector<32xi1>
-    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 2 : i64}>
-        : i64, vector<32xindex>, vector<32xi1> -> vector<32x2xf32>
-    gpu.return %v : vector<32x2xf32>
-  }
-}
-
 // -----
 // Store with vector.step offsets: hint stamped with factor = 2 on the store
 // op.
@@ -99,7 +80,7 @@ gpu.module @kernel_explicit_chunk [#xevm.target<chip = "pvc">] {
 func.func @store_step_hint(%ptr: i64, %v: vector<32xf32>) {
   %offsets = vector.step : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
-  xegpu.store %v, %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  xegpu.store %v, %ptr[%offsets], %mask
       : vector<32xf32>, i64, vector<32xindex>, vector<32xi1>
   return
 }
diff --git a/mlir/test/Dialect/XeGPU/coalesce-gather-scatter.mlir b/mlir/test/Dialect/XeGPU/coalesce-gather-scatter.mlir
index 171879f18ad87..25237df7f8014 100644
--- a/mlir/test/Dialect/XeGPU/coalesce-gather-scatter.mlir
+++ b/mlir/test/Dialect/XeGPU/coalesce-gather-scatter.mlir
@@ -5,19 +5,16 @@
 // vector.step offsets -> stride 1, fully coalescible.
 // 32 lanes, subgroup_size = 16 (default) -> lane_layout = 16, perLane = 2,
 // max-chunk-size = 8 -> bound = min(32, 8, 2) = 2 -> lane_data = 2.
-// The trivial `chunk_size = 1` attribute is dropped on success since the
-// new lane_data FCD > 1 supersedes it.
 // CHECK-LABEL: func.func @load_step_offsets(
 // CHECK:   %[[STEP:.*]] = vector.step : vector<32xindex>
 // CHECK:   %[[LOAD:.*]] = xegpu.load
-// CHECK-NOT: chunk_size
 // CHECK-SAME: layout = #xegpu.layout<inst_data = [32], lane_layout = [16], lane_data = [2]>
 // CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
 // CHECK:   return %[[LOAD]]
 func.func @load_step_offsets(%ptr: i64) -> vector<32xf32> {
   %offsets = vector.step : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
-  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %ptr[%offsets], %mask
       : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
   return %v : vector<32xf32>
 }
@@ -33,7 +30,7 @@ func.func @load_step_offsets(%ptr: i64) -> vector<32xf32> {
 func.func @load_step_offsets_chunk4(%ptr: i64) -> vector<32xf32> {
   %offsets = vector.step : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
-  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %ptr[%offsets], %mask
       : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
   return %v : vector<32xf32>
 }
@@ -50,7 +47,7 @@ func.func @load_dense_ap_offsets(%ptr: i64) -> vector<32xi32> {
     16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
   ]> : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
-  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %ptr[%offsets], %mask
       : i64, vector<32xindex>, vector<32xi1> -> vector<32xi32>
   return %v : vector<32xi32>
 }
@@ -67,7 +64,7 @@ func.func @load_stride4_unchanged(%ptr: i64) -> vector<32xf32> {
   %splat = vector.broadcast %c4 : index to vector<32xindex>
   %offsets = arith.muli %step, %splat : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
-  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %ptr[%offsets], %mask
       : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
   return %v : vector<32xf32>
 }
@@ -80,7 +77,7 @@ func.func @load_stride4_unchanged(%ptr: i64) -> vector<32xf32> {
 // CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
 func.func @load_partial_mask_unchanged(%ptr: i64, %mask: vector<32xi1>) -> vector<32xf32> {
   %offsets = vector.step : vector<32xindex>
-  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %ptr[%offsets], %mask
       : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
   return %v : vector<32xf32>
 }
@@ -94,7 +91,7 @@ func.func @load_partial_mask_unchanged(%ptr: i64, %mask: vector<32xi1>) -> vecto
 func.func @store_step_offsets(%ptr: i64, %v: vector<32xf32>) {
   %offsets = vector.step : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
-  xegpu.store %v, %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  xegpu.store %v, %ptr[%offsets], %mask
       : vector<32xf32>, i64, vector<32xindex>, vector<32xi1>
   return
 }
@@ -108,7 +105,7 @@ func.func @store_step_offsets(%ptr: i64, %v: vector<32xf32>) {
 func.func @store_broadcast_offsets_unchanged(%ptr: i64, %v: vector<32xf32>) {
   %offsets = arith.constant dense<0> : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
-  xegpu.store %v, %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  xegpu.store %v, %ptr[%offsets], %mask
       : vector<32xf32>, i64, vector<32xindex>, vector<32xi1>
   return
 }
@@ -122,7 +119,7 @@ func.func @store_broadcast_offsets_unchanged(%ptr: i64, %v: vector<32xf32>) {
 func.func @load_memref_step(%m: memref<1024xf32>) -> vector<32xf32> {
   %offsets = vector.step : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
-  %v = xegpu.load %m[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %m[%offsets], %mask
       : memref<1024xf32>, vector<32xindex>, vector<32xi1> -> vector<32xf32>
   return %v : vector<32xf32>
 }
@@ -139,7 +136,7 @@ func.func @load_2d_leading_unit(%ptr: i64) -> vector<1x32xf32> {
   %step = vector.step : vector<32xindex>
   %offsets = vector.shape_cast %step : vector<32xindex> to vector<1x32xindex>
   %mask = arith.constant dense<true> : vector<1x32xi1>
-  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %ptr[%offsets], %mask
       : i64, vector<1x32xindex>, vector<1x32xi1> -> vector<1x32xf32>
   return %v : vector<1x32xf32>
 }
@@ -159,7 +156,7 @@ func.func @load_2d_dense_ap(%ptr: i64) -> vector<2x16xf32> {
     [16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]
   ]> : vector<2x16xindex>
   %mask = arith.constant dense<true> : vector<2x16xi1>
-  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %ptr[%offsets], %mask
       : i64, vector<2x16xindex>, vector<2x16xi1> -> vector<2x16xf32>
   return %v : vector<2x16xf32>
 }
@@ -168,7 +165,7 @@ func.func @load_2d_dense_ap(%ptr: i64) -> vector<2x16xf32> {
 // True 2-D dense AP with inner > subgroup_size: each row stride-1 across 32
 // lanes. With subgroup_size = 16 (from #xevm.target chip = "pvc"),
 // lane_layout[inner] = 16, perLane = 32 / 16 = 2, contiguity[inner] = 32,
-// budget = max-chunk-size / 1 = 8. bound = min(32, 8, 2) = 2 -> lane_data = 2.
+// budget = max-chunk-size = 8. bound = min(32, 8, 2) = 2 -> lane_data = 2.
 // CHECK-LABEL: gpu.func @load_2d_dense_ap_wide(
 // CHECK: xegpu.load
 // CHECK-SAME: layout = #xegpu.layout<inst_data = [1, 32], lane_layout = [1, 16], lane_data = [1, 2]>
@@ -182,7 +179,7 @@ gpu.module @kernel [#xevm.target<chip = "pvc">] {
        48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63]
     ]> : vector<2x32xindex>
     %mask = arith.constant dense<true> : vector<2x32xi1>
-    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+    %v = xegpu.load %ptr[%offsets], %mask
         : i64, vector<2x32xindex>, vector<2x32xi1> -> vector<2x32xf32>
     gpu.return %v : vector<2x32xf32>
   }
@@ -202,7 +199,7 @@ gpu.module @kernel_2x32 [#xevm.target<chip = "pvc">] {
     %step = vector.step : vector<64xindex>
     %offsets = vector.shape_cast %step : vector<64xindex> to vector<2x32xindex>
     %mask = arith.constant dense<true> : vector<2x32xi1>
-    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+    %v = xegpu.load %ptr[%offsets], %mask
         : i64, vector<2x32xindex>, vector<2x32xi1> -> vector<2x32xf32>
     gpu.return %v : vector<2x32xf32>
   }
@@ -221,7 +218,7 @@ func.func @load_2d_non_ap_unchanged(%ptr: i64) -> vector<2x16xf32> {
     [16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]
   ]> : vector<2x16xindex>
   %mask = arith.constant dense<true> : vector<2x16xi1>
-  %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  %v = xegpu.load %ptr[%offsets], %mask
       : i64, vector<2x16xindex>, vector<2x16xi1> -> vector<2x16xf32>
   return %v : vector<2x16xf32>
 }
@@ -255,7 +252,7 @@ gpu.module @kernel_reduction [#xevm.target<chip = "pvc">] {
             : vector<32xindex> to vector<2x32xindex>
     %off    = arith.addi %rowsT, %cols2 : vector<2x32xindex>
     %mask   = arith.constant dense<true> : vector<2x32xi1>
-    %v = xegpu.load %ptr[%off], %mask <{chunk_size = 1 : i64}>
+    %v = xegpu.load %ptr[%off], %mask
         : i64, vector<2x32xindex>, vector<2x32xi1> -> vector<2x32xf32>
     gpu.return %v : vector<2x32xf32>
   }
@@ -279,7 +276,7 @@ gpu.module @kernel_divui [#xevm.target<chip = "pvc">] {
     %c2   = arith.constant dense<2> : vector<32xindex>
     %offsets = arith.divui %even, %c2 : vector<32xindex>
     %mask = arith.constant dense<true> : vector<32xi1>
-    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+    %v = xegpu.load %ptr[%offsets], %mask
         : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
     gpu.return %v : vector<32xf32>
   }
@@ -301,7 +298,7 @@ gpu.module @kernel_divsi [#xevm.target<chip = "pvc">] {
     %c2   = arith.constant dense<2> : vector<32xindex>
     %offsets = arith.divsi %even, %c2 : vector<32xindex>
     %mask = arith.constant dense<true> : vector<32xi1>
-    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+    %v = xegpu.load %ptr[%offsets], %mask
         : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
     gpu.return %v : vector<32xf32>
   }
@@ -311,7 +308,7 @@ gpu.module @kernel_divsi [#xevm.target<chip = "pvc">] {
 // `remui` by a constant that divides the inner stride: every element of a
 // row is the same residue class -> inner-dim constant. The decision picks
 // the broadcast case, which is currently disabled, so the load is left
-// alone (no layout, no chunk_size change).
+// alone (no layout change).
 // CHECK-LABEL: gpu.func @load_remui_inner_uniform(
 // CHECK: xegpu.load
 // CHECK-NOT: lane_data
@@ -325,7 +322,7 @@ gpu.module @kernel_remui [#xevm.target<chip = "pvc">] {
     // (even % 2) is uniformly 0 along the inner dim.
     %offsets = arith.remui %even, %c2 : vector<16xindex>
     %mask = arith.constant dense<true> : vector<16xi1>
-    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+    %v = xegpu.load %ptr[%offsets], %mask
         : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
     gpu.return %v : vector<16xf32>
   }
@@ -347,7 +344,7 @@ gpu.module @kernel_andi [#xevm.target<chip = "pvc">] {
     %m = arith.constant dense<1> : vector<16xindex>
     %offsets = arith.andi %even, %m : vector<16xindex>
     %mask = arith.constant dense<true> : vector<16xi1>
-    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+    %v = xegpu.load %ptr[%offsets], %mask
         : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
     gpu.return %v : vector<16xf32>
   }
@@ -367,7 +364,7 @@ gpu.module @kernel_shli [#xevm.target<chip = "pvc">] {
     %k = arith.constant dense<1> : vector<32xindex>
     %offsets = arith.shli %step, %k : vector<32xindex>
     %mask = arith.constant dense<true> : vector<32xi1>
-    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+    %v = xegpu.load %ptr[%offsets], %mask
         : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
     gpu.return %v : vector<32xf32>
   }
@@ -388,7 +385,7 @@ gpu.module @kernel_shli_shrui [#xevm.target<chip = "pvc">] {
     %doubled = arith.shli %step, %k : vector<32xindex>
     %offsets = arith.shrui %doubled, %k : vector<32xindex>
     %mask = arith.constant dense<true> : vector<32xi1>
-    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+    %v = xegpu.load %ptr[%offsets], %mask
         : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
     gpu.return %v : vector<32xf32>
   }
@@ -412,7 +409,7 @@ gpu.module @kernel_select [#xevm.target<chip = "pvc">] {
     %baseB = arith.addi %step, %b : vector<32xindex>
     %offsets = arith.select %cond, %baseA, %baseB : vector<32xindex>
     %mask = arith.constant dense<true> : vector<32xi1>
-    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+    %v = xegpu.load %ptr[%offsets], %mask
         : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
     gpu.return %v : vector<32xf32>
   }
@@ -433,7 +430,7 @@ gpu.module @kernel_divui_neg [#xevm.target<chip = "pvc">] {
     %c3   = arith.constant dense<3> : vector<16xindex>
     %offsets = arith.divui %even, %c3 : vector<16xindex>
     %mask = arith.constant dense<true> : vector<16xi1>
-    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+    %v = xegpu.load %ptr[%offsets], %mask
         : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
     gpu.return %v : vector<16xf32>
   }
@@ -449,28 +446,7 @@ func.func @store_2d_step(%ptr: i64, %v: vector<1x32xf32>) {
   %step = vector.step : vector<32xindex>
   %offsets = vector.shape_cast %step : vector<32xindex> to vector<1x32xindex>
   %mask = arith.constant dense<true> : vector<1x32xi1>
-  xegpu.store %v, %ptr[%offsets], %mask <{chunk_size = 1 : i64}>
+  xegpu.store %v, %ptr[%offsets], %mask
       : vector<1x32xf32>, i64, vector<1x32xindex>, vector<1x32xi1>
   return
 }
-
-// -----
-// An op that already declares chunk_size = 2 is left alone: the verifier
-// requires a particular value/mask shape relationship for chunked ops, so
-// the pass conservatively skips when an explicit chunk_size > 1 is set
-// (a downstream pass has already committed to that per-lane chunked
-// access).
-// CHECK-LABEL: gpu.func @load_explicit_chunk_unchanged(
-// CHECK: xegpu.load
-// CHECK-NOT: lane_data
-// CHECK-SAME: <{chunk_size = 2 : i64}>
-// CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32x2xf32>
-gpu.module @kernel_explicit_chunk [#xevm.target<chip = "pvc">] {
-  gpu.func @load_explicit_chunk_unchanged(%ptr: i64) -> vector<32x2xf32> {
-    %offsets = vector.step : vector<32xindex>
-    %mask = arith.constant dense<true> : vector<32xi1>
-    %v = xegpu.load %ptr[%offsets], %mask <{chunk_size = 2 : i64}>
-        : i64, vector<32xindex>, vector<32xi1> -> vector<32x2xf32>
-    gpu.return %v : vector<32x2xf32>
-  }
-}
diff --git a/mlir/test/lib/Dialect/XeGPU/TestXeGPUTransforms.cpp b/mlir/test/lib/Dialect/XeGPU/TestXeGPUTransforms.cpp
index b71bd6a0be618..f24b4dccaab04 100644
--- a/mlir/test/lib/Dialect/XeGPU/TestXeGPUTransforms.cpp
+++ b/mlir/test/lib/Dialect/XeGPU/TestXeGPUTransforms.cpp
@@ -484,7 +484,7 @@ struct TestXeGPUCoalesceGatherScatter
     options.maxChunkSize = maxChunkSize;
     xegpu::runCoalesceGatherScatterAnalysis(getOperation(), options);
     if (!analyzeOnly)
-      xegpu::applyCoalesceGatherScatterHints(getOperation());
+      xegpu::coalesceGatherScatter(getOperation());
   }
 };
 

>From fbaaaf14c113fbfa0b18f7907fd87c3c5483cea0 Mon Sep 17 00:00:00 2001
From: "Shahneous Bari, Md Abdullah" <md.abdullah.shahneous.bari at intel.com>
Date: Wed, 10 Jun 2026 20:25:51 +0000
Subject: [PATCH 3/7] Fix clang-format issue.

---
 .../Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp   | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
index 52f1b321be900..788a5774afc91 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
@@ -1077,8 +1077,8 @@ static void analyzeAndStampHint(OpTy op, DataFlowSolver &solver,
   const auto *lat = solver.lookupState<AxisInfoLattice>(op.getOffsets());
   if (!lat || !lat->getValue().isInitialized())
     return;
-  auto d = decide(lat->getValue(), offsetsTy.getShape(), maxChunkSize,
-                  subgroupSize);
+  auto d =
+      decide(lat->getValue(), offsetsTy.getShape(), maxChunkSize, subgroupSize);
   if (d.kind != CoalesceDecision::Kind::Chunked)
     return;
   auto hint = xegpu::CoalesceHintAttr::get(op.getContext(), d.factor);

>From 46cdcd81726941549fdc5ce87b29a1a76fa23576 Mon Sep 17 00:00:00 2001
From: "Shahneous Bari, Md Abdullah" <md.abdullah.shahneous.bari at intel.com>
Date: Thu, 18 Jun 2026 15:47:20 +0000
Subject: [PATCH 4/7] [mlir][XeGPU] Make coalesce_hint an optional op attribute
 (user-authorable).

Promotes `coalesce_hint` from a loose discardable attribute to an inherent
optional attribute of `xegpu.load` / `xegpu.store` (`OptionalAttr<
XeGPU_CoalesceHintAttr>`). This gives a second source for the hint: a user
(or a higher-level frontend that already knows the access is
contiguous-per-lane) can set it directly, in which case the
coalesce-gather-scatter analysis is bypassed for that op.

  - The analysis now yields to a pre-existing hint: analyzeAndStampHint skips
    an op that already carries `coalesce_hint` (user-set, or stamped by an
    earlier propagation level). This also makes it idempotent across the lane
    and inst-data runs. It uses the generated typed accessors
    (get/set/removeCoalesceHintAttr) instead of string-keyed discardable-attr
    access; getCoalesceHintAttrName() is removed.
  - Verification: the attribute verifier keeps `factor` a power of two >= 2;
    a new op-level check (isValidCoalesceHint) requires the innermost offsets
    dimension to be a multiple of `factor`. The `lane_layout * factor` split
    is chip-dependent and is still checked at apply time.
  - In-tree positional builders/creators of these ops pass the new trailing
    optional attribute explicitly.

Tests: ops.mlir round-trips a user-set coalesce_hint on load and store;
invalid.mlir covers the three rejections (non-power-of-two, < 2, non-dividing
factor); coalesce-gather-scatter-analyze.mlir adds a user-hint-preserved case.

Co-Authored-By: Claude Opus 4.8 <noreply at anthropic.com>
---
 .../mlir/Dialect/XeGPU/IR/XeGPUAttrs.td       | 36 ++++++++++--------
 .../include/mlir/Dialect/XeGPU/IR/XeGPUOps.td |  6 ++-
 .../Dialect/XeGPU/Transforms/Transforms.h     |  5 ---
 .../VectorToXeGPU/VectorToXeGPU.cpp           |  8 ++--
 mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp        | 37 +++++++++++++++++--
 .../Transforms/XeGPUCoalesceGatherScatter.cpp | 33 +++++++++--------
 .../Transforms/XeGPUSgToLaneDistribute.cpp    |  5 ++-
 .../Dialect/XeGPU/Transforms/XeGPUUnroll.cpp  |  6 ++-
 .../Transforms/XeGPUWgToSgDistribute.cpp      |  5 ++-
 .../coalesce-gather-scatter-analyze.mlir      | 22 +++++++++--
 mlir/test/Dialect/XeGPU/invalid.mlir          | 24 ++++++++++++
 mlir/test/Dialect/XeGPU/ops.mlir              | 17 +++++++++
 12 files changed, 148 insertions(+), 56 deletions(-)

diff --git a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
index 292302a74934c..e16ee5031785c 100644
--- a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
+++ b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
@@ -928,27 +928,31 @@ def XeGPU_MemLayoutAttr : XeGPUAttr<"MemLayout", "mem_layout"> {
 }
 
 def XeGPU_CoalesceHintAttr : XeGPUAttr<"CoalesceHint", "coalesce_hint"> {
-  let summary = [{Per-op hint stamped by the coalesce-gather-scatter analysis.}];
+  let summary = [{Optional request to coalesce a gather/scatter access.}];
 
   let description = [{
-    `CoalesceHintAttr` is a discardable attribute attached by the
-    coalesce-gather-scatter analysis to `xegpu.load` / `xegpu.store` ops
-    whose offsets describe a coalescible access pattern. `factor` is the
-    number of contiguous elements grouped together along the innermost
-    (fastest-changing) dimension of the access.
-
-    The attribute is purely advisory: it records the analysis's decision
-    without committing to a layout. A consumer turns `factor` into a concrete
-    `xegpu.layout` (deriving the `lane_layout` / `lane_data` split from the
-    op's offsets inner extent and the chip's subgroup size) and removes the
-    hint. A consumer that decides not to act on the hint (for example, a
-    propagator that detects a conflict with an anchor-driven layout) should
-    likewise remove the attribute rather than leave it dangling.
+    `CoalesceHintAttr` is the optional `coalesce_hint` attribute of
+    `xegpu.load` / `xegpu.store`. It requests that the access group `factor`
+    contiguous elements per lane along the innermost (fastest-changing)
+    dimension of the offsets. `factor` must be a power of two `>= 2`, and the
+    innermost offsets dimension must be a multiple of it.
+
+    The hint has two sources:
+      - it can be set directly by the user (or a higher-level frontend that
+        already knows the access is contiguous-per-lane), in which case the
+        coalesce-gather-scatter analysis is not consulted for that op; or
+      - it is stamped by the analysis when it proves the access coalescible.
+
+    The hint is a *request*, not a commitment: a consumer turns `factor` into
+    a concrete `xegpu.layout` (deriving the `lane_layout` / `lane_data` split
+    from the op's offsets inner extent and the chip's subgroup size) and
+    clears the hint. A consumer that declines (for example, layout propagation
+    detecting a conflict with an anchor-driven layout) likewise clears it
+    rather than leaving it dangling.
 
     Example:
     ```mlir
-    %v = xegpu.load %ptr[%offsets], %mask
-            {xegpu.coalesce_hint = #xegpu.coalesce_hint<factor = 4>}
+    %v = xegpu.load %ptr[%offsets], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 4>}>
             : i64, vector<64xindex>, vector<64xi1> -> vector<64xf32>
     ```
   }];
diff --git a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td
index af8c30742bcb2..a58ee7430a56d 100644
--- a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td
+++ b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td
@@ -773,7 +773,8 @@ def XeGPU_LoadGatherOp : XeGPU_Op<"load", [MemoryEffects<[MemRead]>, AnchorLayou
       OptionalAttr<XeGPU_CacheHintAttr>:$l1_hint,
       OptionalAttr<XeGPU_CacheHintAttr>:$l2_hint,
       OptionalAttr<XeGPU_CacheHintAttr>:$l3_hint,
-      OptionalAttr<DistributeLayoutAttr>:$layout);
+      OptionalAttr<DistributeLayoutAttr>:$layout,
+      OptionalAttr<XeGPU_CoalesceHintAttr>:$coalesce_hint);
   let results = (outs XeGPU_ValueOrScalarType:$value);
 
   let extraClassDeclaration = extraBaseClassDeclaration # [{
@@ -903,7 +904,8 @@ def XeGPU_StoreScatterOp : XeGPU_Op<"store", [MemoryEffects<[MemWrite]>, AnchorL
       OptionalAttr<XeGPU_CacheHintAttr>:$l1_hint,
       OptionalAttr<XeGPU_CacheHintAttr>:$l2_hint,
       OptionalAttr<XeGPU_CacheHintAttr>:$l3_hint,
-      OptionalAttr<DistributeLayoutAttr>:$layout);
+      OptionalAttr<DistributeLayoutAttr>:$layout,
+      OptionalAttr<XeGPU_CoalesceHintAttr>:$coalesce_hint);
 
   let extraClassDeclaration = extraBaseClassDeclaration#[{
     Type getDestType() {
diff --git a/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h b/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h
index fb560acc3dc48..e49b95cee05fa 100644
--- a/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h
+++ b/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h
@@ -84,11 +84,6 @@ void populateXeGPUSgToLaneDistributeTypeConversionAndLegality(
 // Coalesce gather/scatter analysis + apply.
 //===----------------------------------------------------------------------===//
 
-/// Discardable attribute name carrying the coalesce hint
-/// (`#xegpu.coalesce_hint<factor = N>`) stamped by
-/// `runCoalesceGatherScatterAnalysis`.
-inline StringRef getCoalesceHintAttrName() { return "xegpu.coalesce_hint"; }
-
 /// Options controlling `runCoalesceGatherScatterAnalysis`.
 struct CoalesceGatherScatterAnalysisOptions {
   /// Upper bound on the number of contiguous elements grouped per lane by
diff --git a/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp b/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp
index d8eca95cbf23c..150ff0ccb8ade 100644
--- a/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp
+++ b/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp
@@ -498,7 +498,7 @@ static LogicalResult lowerToScatteredLoadOp(vector::TransferReadOp readOp,
       /*l1_hint=*/xegpu::CachePolicyAttr{},
       /*l2_hint=*/xegpu::CachePolicyAttr{},
       /*l3_hint=*/xegpu::CachePolicyAttr{},
-      /*layout=*/nullptr);
+      /*layout=*/nullptr, /*coalesce_hint=*/nullptr);
 
   rewriter.replaceOp(readOp, gatherOp.getResult());
   return success();
@@ -533,7 +533,7 @@ static LogicalResult lowerToScatteredStoreOp(vector::TransferWriteOp writeOp,
                                 /*l1_hint=*/xegpu::CachePolicyAttr{},
                                 /*l2_hint=*/xegpu::CachePolicyAttr{},
                                 /*l3_hint=*/xegpu::CachePolicyAttr{},
-                                /*layout=*/nullptr);
+                                /*layout=*/nullptr, /*coalesce_hint=*/nullptr);
   rewriter.eraseOp(writeOp);
   return success();
 }
@@ -789,7 +789,7 @@ struct GatherLowering : public OpRewritePattern<vector::GatherOp> {
         /*l1_hint=*/xegpu::CachePolicyAttr{},
         /*l2_hint=*/xegpu::CachePolicyAttr{},
         /*l3_hint=*/xegpu::CachePolicyAttr{},
-        /*layout=*/nullptr);
+        /*layout=*/nullptr, /*coalesce_hint=*/nullptr);
 
     auto selectOp =
         arith::SelectOp::create(rewriter, loc, gatherOp.getMask(),
@@ -824,7 +824,7 @@ struct ScatterLowering : public OpRewritePattern<vector::ScatterOp> {
                                   /*l1_hint=*/xegpu::CachePolicyAttr{},
                                   /*l2_hint=*/xegpu::CachePolicyAttr{},
                                   /*l3_hint=*/xegpu::CachePolicyAttr{},
-                                  /*layout=*/nullptr);
+                                  /*layout=*/nullptr, /*coalesce_hint=*/nullptr);
     rewriter.eraseOp(scatterOp);
     return success();
   }
diff --git a/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp b/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
index 64be7fca5d40f..ef98ae84523a0 100644
--- a/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
+++ b/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
@@ -113,6 +113,29 @@ isValidGatherScatterBufferParams(Type offsetsTy, Type maskTy,
   return success();
 }
 
+// Validates a user-provided (or analysis-stamped) `coalesce_hint` against the
+// op's offsets type. The hint requests grouping `factor` contiguous elements
+// per lane along the innermost (fastest-changing) dimension, so that dimension
+// must be a multiple of `factor`. The further `lane_layout * factor` split is
+// chip-dependent (subgroup size) and is checked when the hint is lowered to a
+// layout, not here.
+static LogicalResult
+isValidCoalesceHint(xegpu::CoalesceHintAttr hint, Type offsetsTy,
+                    function_ref<InFlightDiagnostic()> emitError) {
+  if (!hint)
+    return success();
+  auto offsetsVecTy = dyn_cast<VectorType>(offsetsTy);
+  if (!offsetsVecTy)
+    return emitError()
+           << "coalesce_hint requires vector offsets (one per lane).";
+  int64_t factor = hint.getFactor().getInt();
+  int64_t inner = offsetsVecTy.getShape().back();
+  if (inner % factor != 0)
+    return emitError() << "coalesce_hint factor " << factor
+                       << " must divide the innermost offsets dim " << inner;
+  return success();
+}
+
 LogicalResult
 IsValidMatrixOpParams(VectorType dataTy, MemDescType mdescTy,
                       UnitAttr subgroup_block_io, DistributeLayoutAttr layout,
@@ -582,6 +605,9 @@ LogicalResult LoadGatherOp::verify() {
   }
 
   auto offsetsTy = getOffsets().getType();
+  if (failed(isValidCoalesceHint(getCoalesceHintAttr(), offsetsTy,
+                                 [&]() { return emitOpError(); })))
+    return failure();
   return isValidGatherScatterBufferParams(offsetsTy, maskTy, valueTy, chunkSize,
                                           [&]() { return emitOpError(); });
 }
@@ -599,7 +625,7 @@ void LoadGatherOp::build(OpBuilder &builder, OperationState &state,
   auto offset = vector::FromElementsOp::create(builder, loc, type, values);
 
   build(builder, state, valueType, source, offset, mask, chunk_size, l1_hint,
-        l2_hint, l3_hint, /*anchor_layout=*/nullptr);
+        l2_hint, l3_hint, /*anchor_layout=*/nullptr, /*coalesce_hint=*/nullptr);
 }
 
 void LoadGatherOp::build(OpBuilder &builder, OperationState &state,
@@ -616,7 +642,7 @@ void LoadGatherOp::build(OpBuilder &builder, OperationState &state,
   auto offset = vector::FromElementsOp::create(builder, loc, type, values);
 
   build(builder, state, valueType, source, offset, mask, chunk_size, l1_hint,
-        l2_hint, l3_hint, layout);
+        l2_hint, l3_hint, layout, /*coalesce_hint=*/nullptr);
 }
 
 //===----------------------------------------------------------------------===//
@@ -648,6 +674,9 @@ LogicalResult StoreScatterOp::verify() {
   }
 
   auto offsetsTy = getOffsets().getType();
+  if (failed(isValidCoalesceHint(getCoalesceHintAttr(), offsetsTy,
+                                 [&]() { return emitOpError(); })))
+    return failure();
   return isValidGatherScatterBufferParams(offsetsTy, maskTy, valueTy, chunkSize,
                                           [&]() { return emitOpError(); });
 }
@@ -667,7 +696,7 @@ void StoreScatterOp::build(OpBuilder &builder, OperationState &state,
 
   // Call the correct builder overload that does not expect result types.
   build(builder, state, value, dest, offset, mask, chunk_size, l1_hint, l2_hint,
-        l3_hint, /*anchor_layout=*/nullptr);
+        l3_hint, /*anchor_layout=*/nullptr, /*coalesce_hint=*/nullptr);
 }
 
 void StoreScatterOp::build(
@@ -683,7 +712,7 @@ void StoreScatterOp::build(
 
   // Call the correct builder overload that does not expect result types.
   build(builder, state, value, dest, offset, mask, chunk_size, l1_hint, l2_hint,
-        l3_hint, layout);
+        l3_hint, layout, /*coalesce_hint=*/nullptr);
 }
 
 //===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
index 788a5774afc91..89b6c5ac9c7da 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
@@ -1068,6 +1068,13 @@ static void analyzeAndStampHint(OpTy op, DataFlowSolver &solver,
                                 unsigned maxChunkSize) {
   if (!isCandidateForCoalesce(op))
     return;
+  // A user-provided `coalesce_hint` takes precedence: if the op already
+  // carries one (authored by the user, or stamped by an earlier propagation
+  // level), leave it untouched. The analysis is only a fallback that fills in
+  // a hint where the user did not request one. This also makes the analysis
+  // idempotent across the lane and inst-data propagation runs.
+  if (op.getCoalesceHintAttr())
+    return;
   // Do not coalesce accesses tied to a reduction (see isReductionTied);
   // reduction coalescing is added in follow-up PRs.
   if (isReductionTied(op))
@@ -1081,8 +1088,8 @@ static void analyzeAndStampHint(OpTy op, DataFlowSolver &solver,
       decide(lat->getValue(), offsetsTy.getShape(), maxChunkSize, subgroupSize);
   if (d.kind != CoalesceDecision::Kind::Chunked)
     return;
-  auto hint = xegpu::CoalesceHintAttr::get(op.getContext(), d.factor);
-  op->setAttr(xegpu::getCoalesceHintAttrName(), hint);
+  op.setCoalesceHintAttr(
+      xegpu::CoalesceHintAttr::get(op.getContext(), d.factor));
 }
 
 /// Apply a stamped hint on `op`: build a lane_layout/lane_data/inst_data
@@ -1092,8 +1099,7 @@ static void analyzeAndStampHint(OpTy op, DataFlowSolver &solver,
 /// malformed.
 template <typename OpTy>
 static LogicalResult applyHintOnOp(OpTy op) {
-  auto hint = op->template getAttrOfType<xegpu::CoalesceHintAttr>(
-      xegpu::getCoalesceHintAttrName());
+  auto hint = op.getCoalesceHintAttr();
   if (!hint)
     return success(); // no hint: idempotent no-op.
 
@@ -1113,7 +1119,7 @@ static LogicalResult applyHintOnOp(OpTy op) {
   auto layout = buildLaneDataLayout(op.getContext(), valueTy.getRank(),
                                     laneLayout, factor);
   op.setLayoutAttr(layout);
-  op->removeAttr(xegpu::getCoalesceHintAttrName());
+  op.removeCoalesceHintAttr();
   return success();
 }
 
@@ -1126,9 +1132,8 @@ static LogicalResult applyHintOnOp(Operation *op) {
     return applyHintOnOp(load);
   if (auto store = dyn_cast<xegpu::StoreScatterOp>(op))
     return applyHintOnOp(store);
-  // Hint attached to an unsupported op: silently drop it.
-  if (op->hasAttr(xegpu::getCoalesceHintAttrName()))
-    op->removeAttr(xegpu::getCoalesceHintAttrName());
+  // `coalesce_hint` is an inherent attribute of LoadGatherOp / StoreScatterOp
+  // only, so no other op can carry one.
   return success();
 }
 
@@ -1159,16 +1164,14 @@ void mlir::xegpu::runCoalesceGatherScatterAnalysis(
 }
 
 void mlir::xegpu::coalesceGatherScatter(Operation *root) {
-  root->walk([&](Operation *op) {
-    if (op->hasAttr(xegpu::getCoalesceHintAttrName()))
-      (void)applyHintOnOp(op);
-  });
+  root->walk([&](Operation *op) { (void)applyHintOnOp(op); });
 }
 
 void mlir::xegpu::clearCoalesceGatherScatterHints(Operation *root) {
-  StringRef name = xegpu::getCoalesceHintAttrName();
   root->walk([&](Operation *op) {
-    if (op->hasAttr(name))
-      op->removeAttr(name);
+    if (auto load = dyn_cast<xegpu::LoadGatherOp>(op))
+      load.removeCoalesceHintAttr();
+    else if (auto store = dyn_cast<xegpu::StoreScatterOp>(op))
+      store.removeCoalesceHintAttr();
   });
 }
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUSgToLaneDistribute.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUSgToLaneDistribute.cpp
index 8a926754e7cfb..27ff32666a59d 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUSgToLaneDistribute.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUSgToLaneDistribute.cpp
@@ -503,7 +503,7 @@ struct SgToLaneLoadGather : public OpConversionPattern<xegpu::LoadGatherOp> {
     auto newOp = xegpu::LoadGatherOp::create(
         rewriter, op.getLoc(), distResultTy1D, distSource, distOffsets,
         distMask, op.getChunkSizeAttr(), op.getL1HintAttr(), op.getL2HintAttr(),
-        op.getL3HintAttr(), /*layout=*/nullptr);
+        op.getL3HintAttr(), /*layout=*/nullptr, /*coalesce_hint=*/nullptr);
 
     Value result = newOp->getResult(0);
     if (distResultTy1D != distResultTy)
@@ -1034,7 +1034,8 @@ struct SgToLaneStoreScatter
     xegpu::StoreScatterOp::create(rewriter, op.getLoc(), distValue, distDest,
                                   distOffsets, distMask, op.getChunkSizeAttr(),
                                   op.getL1HintAttr(), op.getL2HintAttr(),
-                                  op.getL3HintAttr(), /*layout=*/nullptr);
+                                  op.getL3HintAttr(), /*layout=*/nullptr,
+                                  /*coalesce_hint=*/nullptr);
     rewriter.eraseOp(op);
     return success();
   }
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
index aab36b79845e4..292075b0a8394 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
@@ -566,7 +566,8 @@ struct UnrollLoadGatherOp : public UnrollPattern<xegpu::LoadGatherOp> {
       auto newOp = xegpu::LoadGatherOp::create(
           rewriter, loc, newValueTy, op.getSource(), o, m,
           rewriter.getI64IntegerAttr(chunkSize), op.getL1HintAttr(),
-          op.getL2HintAttr(), op.getL3HintAttr(), layout);
+          op.getL2HintAttr(), op.getL3HintAttr(), layout,
+          /*coalesce_hint=*/nullptr);
       newOps.push_back(newOp);
     }
 
@@ -661,7 +662,8 @@ struct UnrollStoreScatterOp : public UnrollPattern<xegpu::StoreScatterOp> {
       xegpu::StoreScatterOp::create(rewriter, loc, v, op.getDest(), o, m,
                                     rewriter.getI64IntegerAttr(chunkSize),
                                     op.getL1HintAttr(), op.getL2HintAttr(),
-                                    op.getL3HintAttr(), layout);
+                                    op.getL3HintAttr(), layout,
+                                    /*coalesce_hint=*/nullptr);
     }
 
     rewriter.eraseOp(op);
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp
index 1aa03ebc0f376..4f3a1c2a8c947 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp
@@ -901,7 +901,7 @@ struct WgToSgLoadGatherOp : public OpConversionPattern<xegpu::LoadGatherOp> {
       auto newLoadOp = xegpu::LoadGatherOp::create(
           rewriter, loc, newTy, op.getSource(), offsets, mask, chunkSizeAttr,
           op.getL1HintAttr(), op.getL2HintAttr(), op.getL3HintAttr(),
-          newLayout);
+          newLayout, /*coalesce_hint=*/nullptr);
       newLoadOps.push_back(newLoadOp);
     }
     rewriter.replaceOpWithMultiple(op, {newLoadOps});
@@ -947,7 +947,8 @@ struct WgToSgStoreScatterOp
       xegpu::StoreScatterOp::create(rewriter, loc, val, op.getDest(), offs,
                                     mask, chunkSizeAttr, op.getL1HintAttr(),
                                     op.getL2HintAttr(), op.getL3HintAttr(),
-                                    layout.dropSgLayoutAndData());
+                                    layout.dropSgLayoutAndData(),
+                                    /*coalesce_hint=*/nullptr);
     }
     rewriter.eraseOp(op);
     return success();
diff --git a/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir b/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir
index 4b528704607a9..ac1d644dfcd47 100644
--- a/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir
+++ b/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir
@@ -1,7 +1,7 @@
 // RUN: mlir-opt -split-input-file \
 // RUN:   -test-xegpu-coalesce-gather-scatter="analyze-only=true" %s | FileCheck %s
 
-// Analyze-only mode: stamps `xegpu.coalesce_hint` on coalescible ops and
+// Analyze-only mode: stamps the `coalesce_hint` attribute on coalescible ops and
 // leaves the layout unchanged. This test pins the hint attribute contract
 // that the apply API (and downstream propagator integrations) consume.
 
@@ -9,7 +9,7 @@
 // 1-D vector.step, fully coalescible -> hint with factor = 2 stamped.
 // CHECK-LABEL: func.func @load_step_offsets(
 // CHECK: xegpu.load
-// CHECK-SAME: {xegpu.coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}
+// CHECK-SAME: <{coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}>
 // CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
 // CHECK-NOT: lane_data
 func.func @load_step_offsets(%ptr: i64) -> vector<32xf32> {
@@ -24,7 +24,7 @@ func.func @load_step_offsets(%ptr: i64) -> vector<32xf32> {
 // 2-D leading-1 dim: hint stamped on the load with factor = 2.
 // CHECK-LABEL: func.func @load_2d_leading_unit(
 // CHECK: xegpu.load
-// CHECK-SAME: {xegpu.coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}
+// CHECK-SAME: <{coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}>
 // CHECK-SAME: : i64, vector<1x32xindex>, vector<1x32xi1> -> vector<1x32xf32>
 // CHECK-NOT: lane_data
 func.func @load_2d_leading_unit(%ptr: i64) -> vector<1x32xf32> {
@@ -74,7 +74,7 @@ func.func @load_broadcast_offsets_no_hint(%ptr: i64) -> vector<32xf32> {
 // op.
 // CHECK-LABEL: func.func @store_step_hint(
 // CHECK: xegpu.store
-// CHECK-SAME: {xegpu.coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}
+// CHECK-SAME: <{coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}>
 // CHECK-SAME: : vector<32xf32>, i64, vector<32xindex>, vector<32xi1>
 // CHECK-NOT: lane_data
 func.func @store_step_hint(%ptr: i64, %v: vector<32xf32>) {
@@ -84,3 +84,17 @@ func.func @store_step_hint(%ptr: i64, %v: vector<32xf32>) {
       : vector<32xf32>, i64, vector<32xindex>, vector<32xi1>
   return
 }
+
+// -----
+// A user-provided coalesce_hint takes precedence: the analysis must not
+// overwrite it, even if its own heuristic would pick a different factor.
+// CHECK-LABEL: func.func @user_hint_preserved(
+// CHECK: xegpu.load
+// CHECK-SAME: <{coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}>
+func.func @user_hint_preserved(%ptr: i64) -> vector<32xf32> {
+  %offsets = vector.step : vector<32xindex>
+  %mask = arith.constant dense<true> : vector<32xi1>
+  %v = xegpu.load %ptr[%offsets], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 2>}>
+      : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
+  return %v : vector<32xf32>
+}
diff --git a/mlir/test/Dialect/XeGPU/invalid.mlir b/mlir/test/Dialect/XeGPU/invalid.mlir
index 4f45ce574232a..6e7908836043f 100644
--- a/mlir/test/Dialect/XeGPU/invalid.mlir
+++ b/mlir/test/Dialect/XeGPU/invalid.mlir
@@ -772,3 +772,27 @@ func.func @dpas_mx_scale_b_layout_not_distributable(%a : vector<8x16xf8E5M2>, %b
   %1 = xegpu.dpas_mx %a, %b, %acc scale_a = %scale_a_val scale_b = %scale_b_val {layout_b_scale = #layout_b_scale_invalid} : (vector<8x16xf8E5M2>, vector<16x16xf8E5M2>, vector<8x16xf32>, vector<8x2xf8E8M0FNU>, vector<2x16xf8E8M0FNU>) -> vector<8x16xf32>
   return
 }
+
+// -----
+func.func @coalesce_hint_not_power_of_two(%src: i64, %offset: vector<16xindex>, %mask: vector<16xi1>) {
+  // expected-error at +1 {{'factor' : 6 must be a power of two}}
+  %val = xegpu.load %src[%offset], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 6>}>
+      : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
+  return
+}
+
+// -----
+func.func @coalesce_hint_factor_too_small(%src: i64, %offset: vector<16xindex>, %mask: vector<16xi1>) {
+  // expected-error at +1 {{'factor' : 1 must be >= 2}}
+  %val = xegpu.load %src[%offset], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 1>}>
+      : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
+  return
+}
+
+// -----
+func.func @coalesce_hint_factor_does_not_divide(%src: i64, %offset: vector<6xindex>, %mask: vector<6xi1>) {
+  // expected-error at +1 {{coalesce_hint factor 4 must divide the innermost offsets dim 6}}
+  %val = xegpu.load %src[%offset], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 4>}>
+      : i64, vector<6xindex>, vector<6xi1> -> vector<6xf32>
+  return
+}
diff --git a/mlir/test/Dialect/XeGPU/ops.mlir b/mlir/test/Dialect/XeGPU/ops.mlir
index 0b1ac71cdbd32..d0ba3c2119072 100644
--- a/mlir/test/Dialect/XeGPU/ops.mlir
+++ b/mlir/test/Dialect/XeGPU/ops.mlir
@@ -452,6 +452,23 @@ gpu.func @subgroup_store_offset_1(%dest: memref<?xf16>) {
   gpu.return
 }
 
+// CHECK: gpu.func @load_coalesce_hint(%[[arg0:.*]]: i64, %[[arg1:.*]]: vector<16xindex>, %[[arg2:.*]]: vector<16xi1>) {
+gpu.func @load_coalesce_hint(%src: i64, %offset: vector<16xindex>, %mask: vector<16xi1>) {
+  // A user-provided `coalesce_hint` round-trips through the optional op attribute.
+  // CHECK: xegpu.load %[[arg0]][%[[arg1]]], %[[arg2]] <{coalesce_hint = #xegpu.coalesce_hint<factor = 4 : i64>}> : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
+  %val = xegpu.load %src[%offset], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 4>}>
+      : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
+  gpu.return
+}
+
+// CHECK: gpu.func @store_coalesce_hint(%[[arg0:.*]]: vector<16xf32>, %[[arg1:.*]]: i64, %[[arg2:.*]]: vector<16xindex>, %[[arg3:.*]]: vector<16xi1>) {
+gpu.func @store_coalesce_hint(%val: vector<16xf32>, %dest: i64, %offset: vector<16xindex>, %mask: vector<16xi1>) {
+  // CHECK: xegpu.store %[[arg0]], %[[arg1]][%[[arg2]]], %[[arg3]] <{coalesce_hint = #xegpu.coalesce_hint<factor = 4 : i64>}> : vector<16xf32>, i64, vector<16xindex>, vector<16xi1>
+  xegpu.store %val, %dest[%offset], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 4>}>
+      : vector<16xf32>, i64, vector<16xindex>, vector<16xi1>
+  gpu.return
+}
+
 // CHECK: gpu.func @prefetch_offset(%[[arg0:.*]]: ui64) {
 gpu.func @prefetch_offset(%src: ui64) {
   //CHECK: %[[cst:.*]] = arith.constant dense<[0, 8, 16, 24]> : vector<4xindex>

>From f4e39af355551d7229160a843ba6225cbf676d70 Mon Sep 17 00:00:00 2001
From: "Shahneous Bari, Md Abdullah" <md.abdullah.shahneous.bari at intel.com>
Date: Thu, 18 Jun 2026 16:04:25 +0000
Subject: [PATCH 5/7] Fix clang-format.

---
 mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp         | 3 ++-
 mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp | 4 ++--
 2 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp b/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp
index 150ff0ccb8ade..c9cc06614629b 100644
--- a/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp
+++ b/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp
@@ -824,7 +824,8 @@ struct ScatterLowering : public OpRewritePattern<vector::ScatterOp> {
                                   /*l1_hint=*/xegpu::CachePolicyAttr{},
                                   /*l2_hint=*/xegpu::CachePolicyAttr{},
                                   /*l3_hint=*/xegpu::CachePolicyAttr{},
-                                  /*layout=*/nullptr, /*coalesce_hint=*/nullptr);
+                                  /*layout=*/nullptr,
+                                  /*coalesce_hint=*/nullptr);
     rewriter.eraseOp(scatterOp);
     return success();
   }
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp
index 4f3a1c2a8c947..fa1b598767de4 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp
@@ -900,8 +900,8 @@ struct WgToSgLoadGatherOp : public OpConversionPattern<xegpu::LoadGatherOp> {
       auto newLayout = layout.dropSgLayoutAndData();
       auto newLoadOp = xegpu::LoadGatherOp::create(
           rewriter, loc, newTy, op.getSource(), offsets, mask, chunkSizeAttr,
-          op.getL1HintAttr(), op.getL2HintAttr(), op.getL3HintAttr(),
-          newLayout, /*coalesce_hint=*/nullptr);
+          op.getL1HintAttr(), op.getL2HintAttr(), op.getL3HintAttr(), newLayout,
+          /*coalesce_hint=*/nullptr);
       newLoadOps.push_back(newLoadOp);
     }
     rewriter.replaceOpWithMultiple(op, {newLoadOps});

>From 14bf5ffedd3e97b9548de45068c95bda17184cef Mon Sep 17 00:00:00 2001
From: "Shahneous Bari, Md Abdullah" <md.abdullah.shahneous.bari at intel.com>
Date: Wed, 24 Jun 2026 15:14:39 +0000
Subject: [PATCH 6/7] Address review comments.

- rename the analysis file to `XeGPUContiguityAnalysis.cpp`
- keep the file analysis only, and move all mechanism of apply
  to test pass.
- reduce comment clutters.
- replace the coalesce_hint attribute by I64Attr `contiguous_chunk`
---
 .../mlir/Dialect/XeGPU/IR/XeGPUAttrs.td       |  43 --
 .../include/mlir/Dialect/XeGPU/IR/XeGPUOps.td |   4 +-
 .../Dialect/XeGPU/Transforms/Transforms.h     |  34 +-
 .../VectorToXeGPU/VectorToXeGPU.cpp           |   8 +-
 mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp    |  15 -
 mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp        |  45 +-
 ...catter.cpp => XeGPUContiguityAnalysis.cpp} | 404 +++---------------
 .../Transforms/XeGPUSgToLaneDistribute.cpp    |   4 +-
 .../Dialect/XeGPU/Transforms/XeGPUUnroll.cpp  |   4 +-
 .../Transforms/XeGPUWgToSgDistribute.cpp      |   4 +-
 .../coalesce-gather-scatter-analyze.mlir      |  54 ++-
 mlir/test/Dialect/XeGPU/invalid.mlir          |  20 +-
 mlir/test/Dialect/XeGPU/ops.mlir              |  18 +-
 .../lib/Dialect/XeGPU/TestXeGPUTransforms.cpp | 165 ++++++-
 14 files changed, 300 insertions(+), 522 deletions(-)
 rename mlir/lib/Dialect/XeGPU/Transforms/{XeGPUCoalesceGatherScatter.cpp => XeGPUContiguityAnalysis.cpp} (68%)

diff --git a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
index e16ee5031785c..40edce8a60429 100644
--- a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
+++ b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUAttrs.td
@@ -927,49 +927,6 @@ def XeGPU_MemLayoutAttr : XeGPUAttr<"MemLayout", "mem_layout"> {
 
 }
 
-def XeGPU_CoalesceHintAttr : XeGPUAttr<"CoalesceHint", "coalesce_hint"> {
-  let summary = [{Optional request to coalesce a gather/scatter access.}];
-
-  let description = [{
-    `CoalesceHintAttr` is the optional `coalesce_hint` attribute of
-    `xegpu.load` / `xegpu.store`. It requests that the access group `factor`
-    contiguous elements per lane along the innermost (fastest-changing)
-    dimension of the offsets. `factor` must be a power of two `>= 2`, and the
-    innermost offsets dimension must be a multiple of it.
-
-    The hint has two sources:
-      - it can be set directly by the user (or a higher-level frontend that
-        already knows the access is contiguous-per-lane), in which case the
-        coalesce-gather-scatter analysis is not consulted for that op; or
-      - it is stamped by the analysis when it proves the access coalescible.
-
-    The hint is a *request*, not a commitment: a consumer turns `factor` into
-    a concrete `xegpu.layout` (deriving the `lane_layout` / `lane_data` split
-    from the op's offsets inner extent and the chip's subgroup size) and
-    clears the hint. A consumer that declines (for example, layout propagation
-    detecting a conflict with an anchor-driven layout) likewise clears it
-    rather than leaving it dangling.
-
-    Example:
-    ```mlir
-    %v = xegpu.load %ptr[%offsets], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 4>}>
-            : i64, vector<64xindex>, vector<64xi1> -> vector<64xf32>
-    ```
-  }];
-
-  let parameters = (ins "IntegerAttr": $factor);
-
-  let builders = [
-    AttrBuilder<(ins "int64_t":$factor), [{
-      return $_get($_ctxt, IntegerAttr::get(IntegerType::get($_ctxt, 64),
-                                            factor));
-    }]>
-  ];
-
-  let assemblyFormat = "`<` `factor` `=` $factor `>`";
-  let genVerifyDecl = 1;
-}
-
 def AnchorLayoutInterface : OpInterface<"AnchorLayoutInterface"> {
   let cppNamespace = "::mlir::xegpu";
 
diff --git a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td
index a58ee7430a56d..5255448dae33d 100644
--- a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td
+++ b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td
@@ -774,7 +774,7 @@ def XeGPU_LoadGatherOp : XeGPU_Op<"load", [MemoryEffects<[MemRead]>, AnchorLayou
       OptionalAttr<XeGPU_CacheHintAttr>:$l2_hint,
       OptionalAttr<XeGPU_CacheHintAttr>:$l3_hint,
       OptionalAttr<DistributeLayoutAttr>:$layout,
-      OptionalAttr<XeGPU_CoalesceHintAttr>:$coalesce_hint);
+      OptionalAttr<I64Attr>:$contiguous_chunk);
   let results = (outs XeGPU_ValueOrScalarType:$value);
 
   let extraClassDeclaration = extraBaseClassDeclaration # [{
@@ -905,7 +905,7 @@ def XeGPU_StoreScatterOp : XeGPU_Op<"store", [MemoryEffects<[MemWrite]>, AnchorL
       OptionalAttr<XeGPU_CacheHintAttr>:$l2_hint,
       OptionalAttr<XeGPU_CacheHintAttr>:$l3_hint,
       OptionalAttr<DistributeLayoutAttr>:$layout,
-      OptionalAttr<XeGPU_CoalesceHintAttr>:$coalesce_hint);
+      OptionalAttr<I64Attr>:$contiguous_chunk);
 
   let extraClassDeclaration = extraBaseClassDeclaration#[{
     Type getDestType() {
diff --git a/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h b/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h
index e49b95cee05fa..84b4dcfb0a736 100644
--- a/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h
+++ b/mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h
@@ -84,33 +84,13 @@ void populateXeGPUSgToLaneDistributeTypeConversionAndLegality(
 // Coalesce gather/scatter analysis + apply.
 //===----------------------------------------------------------------------===//
 
-/// Options controlling `runCoalesceGatherScatterAnalysis`.
-struct CoalesceGatherScatterAnalysisOptions {
-  /// Upper bound on the number of contiguous elements grouped per lane by
-  /// coalescing. Mirrors the `max-chunk-size` option of the original pass.
-  unsigned maxChunkSize = 8;
-};
-
-/// Run the AxisInfo-based coalescing analysis over `root` and stamp a
-/// `xegpu.coalesce_hint` attribute on every `xegpu.load` / `xegpu.store`
-/// the analysis classifies as coalescible. Ops with an existing non-empty
-/// `lane_data` or a non-uniform mask are skipped (no hint stamped). The hint
-/// is consumed downstream — either by `coalesceGatherScatter` or by a
-/// layout-propagation pass that reads the hint directly.
-void runCoalesceGatherScatterAnalysis(
-    Operation *root, const CoalesceGatherScatterAnalysisOptions &options = {});
-
-/// Walk `root` and turn every stamped `xegpu.coalesce_hint` into an
-/// equivalent `lane_layout` / `lane_data` / `inst_data` layout on its op,
-/// then remove the hint. A hint on an op that cannot be coalesced (e.g. one
-/// that isn't a gather/scatter) is simply dropped. Pairs with
-/// `runCoalesceGatherScatterAnalysis`.
-void coalesceGatherScatter(Operation *root);
-
-/// Walk `root` and remove any leftover `xegpu.coalesce_hint` attributes —
-/// useful as a cleanup after a propagator has decided whether to honor each
-/// hint.
-void clearCoalesceGatherScatterHints(Operation *root);
+/// Run the AxisInfo-based contiguity analysis over `root` and stamp a
+/// `contiguous_chunk` attribute on every `xegpu.load` / `xegpu.store` whose
+/// offsets are contiguous (in runs of >= 2) along the innermost dimension.
+/// The stamped value is the inner-dim contiguity; it is a target-independent
+/// property consumed downstream (e.g. to derive a `lane_data` split). Ops that
+/// already carry a `contiguous_chunk` attribute are left untouched.
+void runCoalesceGatherScatterAnalysis(Operation *root);
 
 /// Collect a set of patterns to unroll xegpu operations to a smaller shapes.
 /// Users can control whether an operation to be unrolled or not, as well as
diff --git a/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp b/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp
index c9cc06614629b..a428a9d82e618 100644
--- a/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp
+++ b/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp
@@ -498,7 +498,7 @@ static LogicalResult lowerToScatteredLoadOp(vector::TransferReadOp readOp,
       /*l1_hint=*/xegpu::CachePolicyAttr{},
       /*l2_hint=*/xegpu::CachePolicyAttr{},
       /*l3_hint=*/xegpu::CachePolicyAttr{},
-      /*layout=*/nullptr, /*coalesce_hint=*/nullptr);
+      /*layout=*/nullptr, /*contiguous_chunk=*/nullptr);
 
   rewriter.replaceOp(readOp, gatherOp.getResult());
   return success();
@@ -533,7 +533,7 @@ static LogicalResult lowerToScatteredStoreOp(vector::TransferWriteOp writeOp,
                                 /*l1_hint=*/xegpu::CachePolicyAttr{},
                                 /*l2_hint=*/xegpu::CachePolicyAttr{},
                                 /*l3_hint=*/xegpu::CachePolicyAttr{},
-                                /*layout=*/nullptr, /*coalesce_hint=*/nullptr);
+                                /*layout=*/nullptr, /*contiguous_chunk=*/nullptr);
   rewriter.eraseOp(writeOp);
   return success();
 }
@@ -789,7 +789,7 @@ struct GatherLowering : public OpRewritePattern<vector::GatherOp> {
         /*l1_hint=*/xegpu::CachePolicyAttr{},
         /*l2_hint=*/xegpu::CachePolicyAttr{},
         /*l3_hint=*/xegpu::CachePolicyAttr{},
-        /*layout=*/nullptr, /*coalesce_hint=*/nullptr);
+        /*layout=*/nullptr, /*contiguous_chunk=*/nullptr);
 
     auto selectOp =
         arith::SelectOp::create(rewriter, loc, gatherOp.getMask(),
@@ -825,7 +825,7 @@ struct ScatterLowering : public OpRewritePattern<vector::ScatterOp> {
                                   /*l2_hint=*/xegpu::CachePolicyAttr{},
                                   /*l3_hint=*/xegpu::CachePolicyAttr{},
                                   /*layout=*/nullptr,
-                                  /*coalesce_hint=*/nullptr);
+                                  /*contiguous_chunk=*/nullptr);
     rewriter.eraseOp(scatterOp);
     return success();
   }
diff --git a/mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp b/mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp
index 8cdef2dd994d8..e92b109c2223e 100644
--- a/mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp
+++ b/mlir/lib/Dialect/XeGPU/IR/XeGPUDialect.cpp
@@ -1213,21 +1213,6 @@ RangeAttr::verify(llvm::function_ref<mlir::InFlightDiagnostic()> emitError,
   return success();
 }
 
-//===----------------------------------------------------------------------===//
-// XeGPU_CoalesceHintAttr
-//===----------------------------------------------------------------------===//
-
-LogicalResult CoalesceHintAttr::verify(
-    llvm::function_ref<mlir::InFlightDiagnostic()> emitError,
-    IntegerAttr factor) {
-  int64_t f = factor.getInt();
-  if (f < 2)
-    return emitError() << "'factor' : " << f << " must be >= 2";
-  if ((f & (f - 1)) != 0)
-    return emitError() << "'factor' : " << f << " must be a power of two";
-  return success();
-}
-
 //===----------------------------------------------------------------------===//
 // XeGPU_TensorDescType
 //===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp b/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
index ef98ae84523a0..1ef90f5915934 100644
--- a/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
+++ b/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
@@ -113,26 +113,26 @@ isValidGatherScatterBufferParams(Type offsetsTy, Type maskTy,
   return success();
 }
 
-// Validates a user-provided (or analysis-stamped) `coalesce_hint` against the
-// op's offsets type. The hint requests grouping `factor` contiguous elements
-// per lane along the innermost (fastest-changing) dimension, so that dimension
-// must be a multiple of `factor`. The further `lane_layout * factor` split is
-// chip-dependent (subgroup size) and is checked when the hint is lowered to a
-// layout, not here.
+// Validates the `contiguous_chunk` attribute against the op's offsets type:
+// the inner offsets dimension is contiguous in runs of `size`, so `size` must
+// be in [2, innermost offsets dim].
 static LogicalResult
-isValidCoalesceHint(xegpu::CoalesceHintAttr hint, Type offsetsTy,
-                    function_ref<InFlightDiagnostic()> emitError) {
-  if (!hint)
+isValidContiguousChunk(std::optional<uint64_t> chunk, Type offsetsTy,
+                       function_ref<InFlightDiagnostic()> emitError) {
+  if (!chunk)
     return success();
   auto offsetsVecTy = dyn_cast<VectorType>(offsetsTy);
   if (!offsetsVecTy)
     return emitError()
-           << "coalesce_hint requires vector offsets (one per lane).";
-  int64_t factor = hint.getFactor().getInt();
+           << "contiguous_chunk requires vector offsets (one per lane).";
+  int64_t size = static_cast<int64_t>(*chunk);
   int64_t inner = offsetsVecTy.getShape().back();
-  if (inner % factor != 0)
-    return emitError() << "coalesce_hint factor " << factor
-                       << " must divide the innermost offsets dim " << inner;
+  if (size < 2)
+    return emitError() << "contiguous_chunk " << size << " must be >= 2";
+  if (size > inner)
+    return emitError() << "contiguous_chunk " << size
+                       << " must not exceed the innermost offsets dim "
+                       << inner;
   return success();
 }
 
@@ -605,8 +605,8 @@ LogicalResult LoadGatherOp::verify() {
   }
 
   auto offsetsTy = getOffsets().getType();
-  if (failed(isValidCoalesceHint(getCoalesceHintAttr(), offsetsTy,
-                                 [&]() { return emitOpError(); })))
+  if (failed(isValidContiguousChunk(getContiguousChunk(), offsetsTy,
+                                    [&]() { return emitOpError(); })))
     return failure();
   return isValidGatherScatterBufferParams(offsetsTy, maskTy, valueTy, chunkSize,
                                           [&]() { return emitOpError(); });
@@ -625,7 +625,8 @@ void LoadGatherOp::build(OpBuilder &builder, OperationState &state,
   auto offset = vector::FromElementsOp::create(builder, loc, type, values);
 
   build(builder, state, valueType, source, offset, mask, chunk_size, l1_hint,
-        l2_hint, l3_hint, /*anchor_layout=*/nullptr, /*coalesce_hint=*/nullptr);
+        l2_hint, l3_hint, /*anchor_layout=*/nullptr,
+        /*contiguous_chunk=*/nullptr);
 }
 
 void LoadGatherOp::build(OpBuilder &builder, OperationState &state,
@@ -642,7 +643,7 @@ void LoadGatherOp::build(OpBuilder &builder, OperationState &state,
   auto offset = vector::FromElementsOp::create(builder, loc, type, values);
 
   build(builder, state, valueType, source, offset, mask, chunk_size, l1_hint,
-        l2_hint, l3_hint, layout, /*coalesce_hint=*/nullptr);
+        l2_hint, l3_hint, layout, /*contiguous_chunk=*/nullptr);
 }
 
 //===----------------------------------------------------------------------===//
@@ -674,8 +675,8 @@ LogicalResult StoreScatterOp::verify() {
   }
 
   auto offsetsTy = getOffsets().getType();
-  if (failed(isValidCoalesceHint(getCoalesceHintAttr(), offsetsTy,
-                                 [&]() { return emitOpError(); })))
+  if (failed(isValidContiguousChunk(getContiguousChunk(), offsetsTy,
+                                    [&]() { return emitOpError(); })))
     return failure();
   return isValidGatherScatterBufferParams(offsetsTy, maskTy, valueTy, chunkSize,
                                           [&]() { return emitOpError(); });
@@ -696,7 +697,7 @@ void StoreScatterOp::build(OpBuilder &builder, OperationState &state,
 
   // Call the correct builder overload that does not expect result types.
   build(builder, state, value, dest, offset, mask, chunk_size, l1_hint, l2_hint,
-        l3_hint, /*anchor_layout=*/nullptr, /*coalesce_hint=*/nullptr);
+        l3_hint, /*anchor_layout=*/nullptr, /*contiguous_chunk=*/nullptr);
 }
 
 void StoreScatterOp::build(
@@ -712,7 +713,7 @@ void StoreScatterOp::build(
 
   // Call the correct builder overload that does not expect result types.
   build(builder, state, value, dest, offset, mask, chunk_size, l1_hint, l2_hint,
-        l3_hint, layout, /*coalesce_hint=*/nullptr);
+        l3_hint, layout, /*contiguous_chunk=*/nullptr);
 }
 
 //===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUContiguityAnalysis.cpp
similarity index 68%
rename from mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
rename to mlir/lib/Dialect/XeGPU/Transforms/XeGPUContiguityAnalysis.cpp
index 89b6c5ac9c7da..1c67503f7445a 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUCoalesceGatherScatter.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUContiguityAnalysis.cpp
@@ -6,31 +6,19 @@
 //
 //===----------------------------------------------------------------------===//
 //
-// This file implements the gather/scatter coalescing analysis. It decides
-// whether an `xegpu.load` / `xegpu.store` gather/scatter accesses `N`
-// contiguous elements per lane along the innermost dimension, and, if so,
-// stamps a `xegpu.coalesce_hint<factor = N>` attribute recording the chosen
-// factor. The decision is driven by a small XeGPU-local axis-info dataflow
-// analysis tracking per-axis `contiguity`, `constancy`, and `divisibility`.
+// This file implements the contiguity analysis. It computes, for a memory
+// operation (i.e., `xegpu.load` / `xegpu.store`), how many elements are
+// contiguous along the innermost offsets dimension, and stamps that count as an
+// attribute on the op. The analysis is a small XeGPU-local
+// axis-info dataflow tracking per-axis `contiguity`, `constancy`, and
+// `divisibility`; the stamped value is the inner-dim `contiguity`.
 //
-// The analysis performs no rewrite. The hint it stamps is turned into a
-// `lane_data` layout by `coalesceGatherScatter` (used by the test pass) or
-// read directly by layout propagation; the actual memory-message rewrite is
-// left to the downstream WG-to-SG / SG-to-Lane distribution passes, which
-// interpret `lane_data`.
+// Contiguity is a target-independent property of the offsets.
 //
 // The analysis tracks per-axis information for vectors of integer / index
-// type at any rank. The coalescing decision is computed against the
-// innermost dimension. 2-D offsets vectors with a leading unit dimension
-// (e.g. `vector<1x32xindex>`) are handled by treating the inner dim as the
-// lane dim.
-//
-// All-equal offsets ("uniform inner dim") are detected by the analysis
-// but the pass currently leaves such ops alone — there is no layout-only
-// encoding for "all lanes load the same scalar", and the previous
-// length-1-load + `vector.broadcast` rewrite was removed because it
-// conflicts with downstream layout propagation.
+// type at any rank, against the innermost dimension.
 //
+// The analysis gets it's inspiration from the Triton Axis info analysis.
 //===----------------------------------------------------------------------===//
 
 #include "mlir/Analysis/DataFlow/DeadCodeAnalysis.h"
@@ -43,14 +31,9 @@
 #include "mlir/Dialect/XeGPU/Transforms/Passes.h"
 #include "mlir/Dialect/XeGPU/Transforms/Transforms.h"
 #include "mlir/Dialect/XeGPU/Utils/XeGPUUtils.h"
-#include "mlir/Dialect/XeGPU/uArch/IntelGpuXe2.h"
 #include "mlir/IR/BuiltinAttributes.h"
 #include "mlir/IR/BuiltinTypes.h"
-#include "mlir/IR/Matchers.h"
-#include "mlir/IR/PatternMatch.h"
-#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
 #include "llvm/ADT/APInt.h"
-#include "llvm/ADT/bit.h"
 #include "llvm/Support/MathExtras.h"
 #include <numeric>
 #include <optional>
@@ -74,25 +57,16 @@ namespace mlir::xegpu::detail::axis_dataflow {
 static constexpr int64_t kAxisInfoTop = 1LL << 30;
 
 /// Per-dimension axis information for an SSA value of integer / index type.
-///   - `contiguity[d]`: the largest N such that consecutive lanes along
-///     dimension `d` differ by exactly 1 in runs of length `N` (lane-stride
-///     1 contiguity).
-///   - `constancy[d]`: the largest N such that consecutive lanes along
-///     dimension `d` are all equal in runs of length `N`.
-///   - `divisibility[d]`: a power-of-two divisor of every element along
-///     dimension `d`.
-///   - `knownConstant`: scalar value if the entire vector is uniformly
-///     known to be a single constant.
-///   - `innerStride`: when set, every "row" along the innermost dimension
-///     is an arithmetic progression with this (constant) step. This is the
-///     general form of which `contiguity`/`constancy` are the two special
-///     cases: `innerStride = 1` is the stride-1 contiguous case (and implies
-///     `contiguity[innerDim] > 1`), `innerStride = 0` is the all-equal case
-///     (implies `constancy[innerDim] > 1`), and any other value (e.g. 4) is a
-///     strided progression that neither `contiguity` nor `constancy` can
-///     represent — both read 1 there. Per-row base may differ across outer
-///     indices; per-row alignment is captured by `divisibility[innerDim]`.
-///     The coalescing decision only acts on the `innerStride = 1` case.
+/// All fields describe runs of consecutive elements along a dimension `d`:
+///   - `contiguity[d]`: longest run that increases by exactly 1.
+///   - `constancy[d]`: longest run of equal values.
+///   - `divisibility[d]`: a power-of-two divisor of every element.
+///   - `knownConstant`: the value, if the whole vector is one constant.
+///   - `innerStride`: if set, each row along the innermost dim is an
+///     arithmetic progression with this step. `1` is the contiguous case,
+///     `0` the all-equal case; any other value is a strided progression that
+///     `contiguity`/`constancy` can't represent (both read 1). Per-row base
+///     may vary; per-row alignment lives in `divisibility[innerDim]`.
 ///
 /// Pessimistic / entry value: contiguity=1, constancy=1, divisibility=1,
 /// innerStride absent.
@@ -586,9 +560,9 @@ class AxisInfoAnalysis
     unsigned r = vt.getRank();
     auto shape = vt.getShape();
     AxisInfo v = AxisInfo::getPessimistic(r);
-    auto unitConstant = [](const AxisInfo &a, unsigned d, int64_t lanes) {
+    auto unitConstant = [](const AxisInfo &a, unsigned d, int64_t extent) {
       return a.knownConstant && *a.knownConstant == 1 &&
-             a.constancy[d] >= lanes;
+             a.constancy[d] >= extent;
     };
     for (unsigned d = 0; d < r; ++d) {
       v.constancy[d] = std::min({shape[d], lhs.constancy[d], rhs.constancy[d]});
@@ -617,22 +591,17 @@ class AxisInfoAnalysis
   }
 
   // arith.divui / arith.divsi / arith.remui / arith.remsi by a uniform
-  // positive constant `c`.
-  //
-  // Division (IsRem=false): when the lhs is an inner-dim AP `(base, s, n)`
-  // with `c | s` and `c | divisibility[inner]` (so the per-row base / c is
-  // exact), the result is an AP with stride `s / c` and inner divisibility
-  // `divisibility[inner] / c`. Special cases: `s/c == 1` flags inner-dim
-  // contiguity; `s/c == 0` flags inner-dim constancy.
+  // positive constant `c`. The lhs must be an arithmetic progression (AP)
+  // along the inner dim, i.e. its values step by a constant stride `s`, and
+  // `c` must divide `s`.
   //
-  // Remainder (IsRem=true): when `c | s`, every element of a row sits at
-  // the same residue class -> inner-dim constant -> `innerStride = 0`,
-  // `constancy[inner] = inner` (matches the analysis's notion of
-  // chunk-uniform values along the inner dim).
+  // Take the inner row `[0, 2, 4, 6, 8, 10, 12, 14]` (stride s = 2) and c = 2:
+  //   - Division `/ 2` gives `[0, 1, 2, 3, 4, 5, 6, 7]`: a new AP with stride
+  //     `s / c = 1`. A resulting stride of 1 is contiguous, 0 is constant.
+  //   - Remainder `% 2` gives `[0, 0, 0, 0, 0, 0, 0, 0]`: every element folds
+  //     to the same residue, so the row is constant (stride 0).
   //
-  // Signed vs unsigned only differs in the constant interpretation; we
-  // require positive constants so the signed/unsigned distinction is moot
-  // here.
+  // We require positive `c`, so signed and unsigned behave the same.
   template <bool IsSigned, bool IsRem, typename OpTy>
   LogicalResult visitDivRem(OpTy op, ArrayRef<const AxisInfoLattice *> operands,
                             ArrayRef<AxisInfoLattice *> results) {
@@ -694,12 +663,16 @@ class AxisInfoAnalysis
     return success();
   }
 
-  // arith.andi: `x & m` with a uniform positive constant mask `m`.
-  // The most useful case is `x & (P - 1)` for `P` a power of 2: this is
-  // equivalent to `x mod P`, so when the lhs is an inner-dim AP with stride
-  // divisible by `P` the result is inner-dim constant. We also handle the
-  // trivial `m == 0` (always zero) and `m == -1`/all-ones (identity)
-  // shapes.
+  // arith.andi: `x & m` with a uniform positive constant mask `m`. The
+  // interesting case is `m = P - 1` for a power of 2 `P`, which is the same
+  // as `x % P` (see visitDivRem): masking an inner row whose stride is a
+  // multiple of `P` folds it to a constant.
+  //
+  // Take the row `[0, 2, 4, 6, 8, 10, 12, 14]` (stride 2) and m = 1 (P = 2):
+  //   `x & 1` gives `[0, 0, 0, 0, 0, 0, 0, 0]`: constant along the inner dim.
+  //
+  // Also handles the trivial masks `m == 0` (always zero) and all-ones
+  // (identity).
   LogicalResult visitAndI(arith::AndIOp op,
                           ArrayRef<const AxisInfoLattice *> operands,
                           ArrayRef<AxisInfoLattice *> results) {
@@ -865,286 +838,40 @@ using ::mlir::xegpu::detail::axis_dataflow::AxisInfo;
 using ::mlir::xegpu::detail::axis_dataflow::AxisInfoLattice;
 
 //===----------------------------------------------------------------------===//
-// Coalescing decision.
+// Analysis driver.
 //===----------------------------------------------------------------------===//
 
-struct CoalesceDecision {
-  enum class Kind { None, Chunked };
-  Kind kind = Kind::None;
-  int64_t laneLayout = 1; // lane_layout along the innermost dim
-  int64_t factor = 1;     // lane_data factor along the innermost dim
-};
-
-/// Largest power-of-two `<= bound` that divides `numLanes`.
-static int64_t largestPow2Divisor(int64_t numLanes, int64_t bound) {
-  if (bound < 2 || numLanes < 2)
-    return 1;
-  int64_t f = std::min<int64_t>(bound, numLanes);
-  // Round down to power of 2.
-  if (!llvm::isPowerOf2_64(f))
-    f = static_cast<int64_t>(llvm::bit_floor(static_cast<uint64_t>(f)));
-  while (f >= 2) {
-    if (numLanes % f == 0)
-      return f;
-    f /= 2;
-  }
-  return 1;
-}
-
-/// Decide how to coalesce given the offsets axis info.
-///
-/// We pick `lane_layout[inner]` first using the same default rule as
-/// `XeGPUPropagateLayout`: `lane_layout[inner] = min(subgroupSize, inner)`,
-/// rounded down to a power-of-2 divisor of `inner`. The remaining lane
-/// budget then becomes `lane_data[inner] = inner / lane_layout`, capped by
-/// `maxChunkSize` and the offsets contiguity.
-static CoalesceDecision decide(const AxisInfo &info,
-                               ArrayRef<int64_t> offsetsShape,
-                               unsigned maxChunkSize, unsigned subgroupSize) {
-  CoalesceDecision d;
-  if (!info.isInitialized() || offsetsShape.empty())
-    return d;
-  unsigned innerDim = offsetsShape.size() - 1;
-  int64_t inner = offsetsShape[innerDim];
-  if (inner < 2)
-    return d;
-
-  // All-equal offsets: every lane sees the same address. There is no
-  // layout-only encoding for this; we leave the op alone.
-  if (info.constancy[innerDim] >= inner)
-    return d;
-
-  // Pick lane_layout first: PropagateLayout's default for a 1-D / inner dim
-  // is `min(subgroupSize, inner)`, rounded down to a divisor of inner.
-  int64_t laneLayout =
-      largestPow2Divisor(inner, std::min<int64_t>(subgroupSize, inner));
-  if (laneLayout < 1)
-    laneLayout = 1;
-
-  // Each lane sees `inner / laneLayout` elements of the offsets vector.
-  int64_t perLane = inner / laneLayout;
-  if (perLane < 2)
-    return d; // already one element per lane, nothing to coalesce.
-
-  int64_t budget = static_cast<int64_t>(maxChunkSize);
-  if (budget < 2)
-    return d;
-
-  int64_t bound = std::min<int64_t>(info.contiguity[innerDim], budget);
-  bound = std::min<int64_t>(bound, perLane);
-  if (bound < 2)
-    return d;
-
-  int64_t factor = largestPow2Divisor(perLane, bound);
-  if (factor < 2)
-    return d;
-  d.kind = CoalesceDecision::Kind::Chunked;
-  d.laneLayout = laneLayout;
-  d.factor = factor;
-  return d;
-}
-
-/// Returns true if `mask` is a constant `dense<true>` vector.
-static bool isAllTrueMask(Value mask) {
-  auto vecTy = dyn_cast<VectorType>(mask.getType());
-  if (!vecTy)
-    return false;
-  auto cst = mask.getDefiningOp<arith::ConstantOp>();
-  if (!cst)
-    return false;
-  auto dense = dyn_cast<DenseIntElementsAttr>(cst.getValue());
-  if (!dense || !dense.isSplat())
-    return false;
-  return dense.getSplatValue<APInt>().getBoolValue();
-}
-
-/// Build a `lane_layout`/`lane_data`/`inst_data` layout of rank `rank`,
-/// with the given lane_layout / lane_data on the innermost dim (1
-/// elsewhere). `inst_data` is `lane_layout * lane_data` per dim, so the
-/// invariant `inst_data[d] == lane_layout[d] * lane_data[d]` holds.
-static xegpu::LayoutAttr buildLaneDataLayout(MLIRContext *ctx, unsigned rank,
-                                             int64_t innerLaneLayout,
-                                             int64_t innerLaneData) {
-  SmallVector<int32_t> laneLayout(rank, 1);
-  SmallVector<int32_t> laneData(rank, 1);
-  SmallVector<int32_t> instData(rank, 1);
-  laneLayout.back() = static_cast<int32_t>(innerLaneLayout);
-  laneData.back() = static_cast<int32_t>(innerLaneData);
-  instData.back() = static_cast<int32_t>(innerLaneLayout * innerLaneData);
-  return xegpu::LayoutAttr::get(ctx, instData, laneLayout, laneData);
-}
-
-//===----------------------------------------------------------------------===//
-// Analysis driver + hint apply.
-//===----------------------------------------------------------------------===//
-
-/// Look up the subgroup size from the enclosing gpu.module's xevm.target.
-/// Falls back to 16 when no target chip is found or the chip is unknown,
-/// matching the typical Intel Xe2 default. This keeps the pass usable on
-/// plain `module { ... }` IR (e.g. unit lit tests) where there's no
-/// gpu.module / xevm.target wrapper.
-static unsigned lookupSubgroupSize(Operation *op) {
-  const auto *uArch =
-      xegpu::uArch::getUArch(xegpu::getChipStr(op).value_or(""));
-  return uArch ? static_cast<unsigned>(uArch->getSubgroupSize()) : 16u;
-}
-
-/// Common analysis preconditions: vector offsets/value, all-true mask,
-/// no existing non-trivial lane_data.
+/// Stamp a `contiguous_chunk` attribute on `op` recording the inner-dim
+/// contiguity computed by the analysis. The contiguity is a target-independent
+/// property of the offsets.
 template <typename OpTy>
-static bool isCandidateForCoalesce(OpTy op) {
+static void analyzeAndStampContiguity(OpTy op, DataFlowSolver &solver) {
   auto offsetsTy = dyn_cast<VectorType>(op.getOffsets().getType());
   if (!offsetsTy || offsetsTy.getNumElements() <= 1)
-    return false;
-  if (!op.getValueType())
-    return false;
-  if (!isAllTrueMask(op.getMask()))
-    return false;
-  if (auto layout = op.getLayoutAttr())
-    if (!layout.getEffectiveLaneDataAsInt().empty())
-      return false;
-  return true;
-}
-
-/// True when `op` (a gather load or scatter store) is tied to a
-/// `vector.multi_reduction`: a load whose result feeds a reduction, or a
-/// store whose stored value comes from one (through layout-neutral /
-/// elementwise / insert glue).
-///
-/// Coalescing such an access is gated off for now: the analysis only sets
-/// `lane_data[FCD]`, and consuming that on a reduction requires reduction-
-/// aware layout handling that is not part of this change. A coalesced store
-/// fed (transitively) by a reduction would seed `lane_data[FCD] = N` that the
-/// reduction result (kept at `lane_data = 1`) cannot match, producing an
-/// unlowerable `xegpu.convert_layout lane_data=[1]<->[N]`. Follow-up PRs that
-/// add reduction coalescing relax this gate.
-template <typename OpTy>
-static bool isReductionTied(OpTy op) {
-  if constexpr (std::is_same_v<OpTy, xegpu::StoreScatterOp>) {
-    // Backward walk of the stored value's slice through the reassembly
-    // (convert_layout / shape_cast / insert(_strided_slice) / elementwise)
-    // that typically sits between a reduction and the store.
-    SmallVector<Value, 8> worklist{op.getValue()};
-    llvm::SmallPtrSet<Operation *, 16> seen;
-    unsigned steps = 0;
-    while (!worklist.empty() && steps++ < 64) {
-      Value v = worklist.pop_back_val();
-      Operation *def = v.getDefiningOp();
-      if (!def || !seen.insert(def).second)
-        continue;
-      if (isa<vector::MultiDimReductionOp>(def))
-        return true;
-      if (isa<vector::ShapeCastOp, vector::BitCastOp, xegpu::ConvertLayoutOp,
-              vector::InsertOp, vector::InsertStridedSliceOp>(def) ||
-          OpTrait::hasElementwiseMappableTraits(def))
-        for (Value operand : def->getOperands())
-          if (isa<VectorType>(operand.getType()))
-            worklist.push_back(operand);
-    }
-    return false;
-  } else {
-    for (Operation *user : op->getUsers()) {
-      Operation *u = user;
-      while (
-          u &&
-          isa<vector::ShapeCastOp, vector::BitCastOp, xegpu::ConvertLayoutOp>(
-              u)) {
-        if (u->getNumResults() != 1 || u->getResult(0).use_empty())
-          break;
-        u = *u->getResult(0).getUsers().begin();
-      }
-      if (u && isa<vector::MultiDimReductionOp>(u))
-        return true;
-    }
-    return false;
-  }
-}
-
-/// Run the analysis on a single op. If the offsets analyze as `Chunked`,
-/// stamp a `xegpu.coalesce_hint` attribute carrying the FCD lane_data
-/// factor.
-template <typename OpTy>
-static void analyzeAndStampHint(OpTy op, DataFlowSolver &solver,
-                                unsigned maxChunkSize) {
-  if (!isCandidateForCoalesce(op))
-    return;
-  // A user-provided `coalesce_hint` takes precedence: if the op already
-  // carries one (authored by the user, or stamped by an earlier propagation
-  // level), leave it untouched. The analysis is only a fallback that fills in
-  // a hint where the user did not request one. This also makes the analysis
-  // idempotent across the lane and inst-data propagation runs.
-  if (op.getCoalesceHintAttr())
     return;
-  // Do not coalesce accesses tied to a reduction (see isReductionTied);
-  // reduction coalescing is added in follow-up PRs.
-  if (isReductionTied(op))
+  // A pre-existing `contiguous_chunk` (user-authored, or stamped by an earlier
+  // run) takes precedence; leave it untouched so the analysis is idempotent.
+  if (op.getContiguousChunk())
     return;
-  auto offsetsTy = cast<VectorType>(op.getOffsets().getType());
-  unsigned subgroupSize = lookupSubgroupSize(op);
   const auto *lat = solver.lookupState<AxisInfoLattice>(op.getOffsets());
   if (!lat || !lat->getValue().isInitialized())
     return;
-  auto d =
-      decide(lat->getValue(), offsetsTy.getShape(), maxChunkSize, subgroupSize);
-  if (d.kind != CoalesceDecision::Kind::Chunked)
+  const AxisInfo &info = lat->getValue();
+  unsigned innerDim = offsetsTy.getRank() - 1;
+  int64_t inner = offsetsTy.getShape()[innerDim];
+  int64_t chunk = std::min<int64_t>(info.contiguity[innerDim], inner);
+  if (chunk < 2)
     return;
-  op.setCoalesceHintAttr(
-      xegpu::CoalesceHintAttr::get(op.getContext(), d.factor));
-}
-
-/// Apply a stamped hint on `op`: build a lane_layout/lane_data/inst_data
-/// layout from the hint's `factor` and the op's offsets inner extent +
-/// chip-derived subgroup size, install it, and remove the hint. Returns
-/// success on apply (or no-op when no hint), failure when the hint is
-/// malformed.
-template <typename OpTy>
-static LogicalResult applyHintOnOp(OpTy op) {
-  auto hint = op.getCoalesceHintAttr();
-  if (!hint)
-    return success(); // no hint: idempotent no-op.
-
-  auto offsetsTy = dyn_cast<VectorType>(op.getOffsets().getType());
-  auto valueTy = op.getValueType();
-  if (!offsetsTy || !valueTy || offsetsTy.getNumElements() <= 1)
-    return failure();
-
-  int64_t factor = hint.getFactor().getInt();
-  int64_t inner = offsetsTy.getShape().back();
-  unsigned subgroupSize = lookupSubgroupSize(op);
-  int64_t laneLayout =
-      largestPow2Divisor(inner, std::min<int64_t>(subgroupSize, inner));
-  if (laneLayout < 1 || inner % (laneLayout * factor) != 0)
-    return failure();
-
-  auto layout = buildLaneDataLayout(op.getContext(), valueTy.getRank(),
-                                    laneLayout, factor);
-  op.setLayoutAttr(layout);
-  op.removeCoalesceHintAttr();
-  return success();
-}
-
-/// Apply a stamped `xegpu.coalesce_hint` on a single op. A hint on an op
-/// that cannot be coalesced (anything other than a gather/scatter) is just
-/// dropped. Returns failure only when the hint is malformed for a real
-/// gather/scatter.
-static LogicalResult applyHintOnOp(Operation *op) {
-  if (auto load = dyn_cast<xegpu::LoadGatherOp>(op))
-    return applyHintOnOp(load);
-  if (auto store = dyn_cast<xegpu::StoreScatterOp>(op))
-    return applyHintOnOp(store);
-  // `coalesce_hint` is an inherent attribute of LoadGatherOp / StoreScatterOp
-  // only, so no other op can carry one.
-  return success();
+  op.setContiguousChunk(chunk);
 }
 
 } // namespace
 
 //===----------------------------------------------------------------------===//
-// Public APIs.
+// Public API.
 //===----------------------------------------------------------------------===//
 
-void mlir::xegpu::runCoalesceGatherScatterAnalysis(
-    Operation *root, const CoalesceGatherScatterAnalysisOptions &options) {
+void mlir::xegpu::runCoalesceGatherScatterAnalysis(Operation *root) {
   DataFlowSolver solver;
   solver.load<dataflow::DeadCodeAnalysis>();
   solver.load<mlir::xegpu::detail::axis_dataflow::AxisInfoAnalysis>();
@@ -1154,24 +881,11 @@ void mlir::xegpu::runCoalesceGatherScatterAnalysis(
   // The solver computed AxisInfo for the whole region in the single
   // `initializeAndRun` above; offsets shared by several gather/scatter ops are
   // analyzed only once. This walk is just per-op point lookups into that
-  // result (no re-analysis), turning each cached fact into a hint.
-  root->walk([&](Operation *op) {
-    if (auto load = dyn_cast<xegpu::LoadGatherOp>(op))
-      analyzeAndStampHint(load, solver, options.maxChunkSize);
-    else if (auto store = dyn_cast<xegpu::StoreScatterOp>(op))
-      analyzeAndStampHint(store, solver, options.maxChunkSize);
-  });
-}
-
-void mlir::xegpu::coalesceGatherScatter(Operation *root) {
-  root->walk([&](Operation *op) { (void)applyHintOnOp(op); });
-}
-
-void mlir::xegpu::clearCoalesceGatherScatterHints(Operation *root) {
+  // result (no re-analysis), turning each cached fact into an attribute.
   root->walk([&](Operation *op) {
     if (auto load = dyn_cast<xegpu::LoadGatherOp>(op))
-      load.removeCoalesceHintAttr();
+      analyzeAndStampContiguity(load, solver);
     else if (auto store = dyn_cast<xegpu::StoreScatterOp>(op))
-      store.removeCoalesceHintAttr();
+      analyzeAndStampContiguity(store, solver);
   });
 }
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUSgToLaneDistribute.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUSgToLaneDistribute.cpp
index 27ff32666a59d..994a7474ceab2 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUSgToLaneDistribute.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUSgToLaneDistribute.cpp
@@ -503,7 +503,7 @@ struct SgToLaneLoadGather : public OpConversionPattern<xegpu::LoadGatherOp> {
     auto newOp = xegpu::LoadGatherOp::create(
         rewriter, op.getLoc(), distResultTy1D, distSource, distOffsets,
         distMask, op.getChunkSizeAttr(), op.getL1HintAttr(), op.getL2HintAttr(),
-        op.getL3HintAttr(), /*layout=*/nullptr, /*coalesce_hint=*/nullptr);
+        op.getL3HintAttr(), /*layout=*/nullptr, /*contiguous_chunk=*/nullptr);
 
     Value result = newOp->getResult(0);
     if (distResultTy1D != distResultTy)
@@ -1035,7 +1035,7 @@ struct SgToLaneStoreScatter
                                   distOffsets, distMask, op.getChunkSizeAttr(),
                                   op.getL1HintAttr(), op.getL2HintAttr(),
                                   op.getL3HintAttr(), /*layout=*/nullptr,
-                                  /*coalesce_hint=*/nullptr);
+                                  /*contiguous_chunk=*/nullptr);
     rewriter.eraseOp(op);
     return success();
   }
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
index 292075b0a8394..558784befe7cd 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
@@ -567,7 +567,7 @@ struct UnrollLoadGatherOp : public UnrollPattern<xegpu::LoadGatherOp> {
           rewriter, loc, newValueTy, op.getSource(), o, m,
           rewriter.getI64IntegerAttr(chunkSize), op.getL1HintAttr(),
           op.getL2HintAttr(), op.getL3HintAttr(), layout,
-          /*coalesce_hint=*/nullptr);
+          /*contiguous_chunk=*/nullptr);
       newOps.push_back(newOp);
     }
 
@@ -663,7 +663,7 @@ struct UnrollStoreScatterOp : public UnrollPattern<xegpu::StoreScatterOp> {
                                     rewriter.getI64IntegerAttr(chunkSize),
                                     op.getL1HintAttr(), op.getL2HintAttr(),
                                     op.getL3HintAttr(), layout,
-                                    /*coalesce_hint=*/nullptr);
+                                    /*contiguous_chunk=*/nullptr);
     }
 
     rewriter.eraseOp(op);
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp
index fa1b598767de4..cd68e7f1808c4 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp
@@ -901,7 +901,7 @@ struct WgToSgLoadGatherOp : public OpConversionPattern<xegpu::LoadGatherOp> {
       auto newLoadOp = xegpu::LoadGatherOp::create(
           rewriter, loc, newTy, op.getSource(), offsets, mask, chunkSizeAttr,
           op.getL1HintAttr(), op.getL2HintAttr(), op.getL3HintAttr(), newLayout,
-          /*coalesce_hint=*/nullptr);
+          /*contiguous_chunk=*/nullptr);
       newLoadOps.push_back(newLoadOp);
     }
     rewriter.replaceOpWithMultiple(op, {newLoadOps});
@@ -948,7 +948,7 @@ struct WgToSgStoreScatterOp
                                     mask, chunkSizeAttr, op.getL1HintAttr(),
                                     op.getL2HintAttr(), op.getL3HintAttr(),
                                     layout.dropSgLayoutAndData(),
-                                    /*coalesce_hint=*/nullptr);
+                                    /*contiguous_chunk=*/nullptr);
     }
     rewriter.eraseOp(op);
     return success();
diff --git a/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir b/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir
index ac1d644dfcd47..8901231e45ff3 100644
--- a/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir
+++ b/mlir/test/Dialect/XeGPU/coalesce-gather-scatter-analyze.mlir
@@ -1,15 +1,15 @@
 // RUN: mlir-opt -split-input-file \
 // RUN:   -test-xegpu-coalesce-gather-scatter="analyze-only=true" %s | FileCheck %s
 
-// Analyze-only mode: stamps the `coalesce_hint` attribute on coalescible ops and
-// leaves the layout unchanged. This test pins the hint attribute contract
-// that the apply API (and downstream propagator integrations) consume.
+// Analyze-only mode: stamps the `contiguous_chunk` attribute on contiguous ops
+// and leaves the layout unchanged. This test pins the attribute contract that
+// the coalescing consumer (and downstream integrations) read.
 
 // -----
-// 1-D vector.step, fully coalescible -> hint with factor = 2 stamped.
+// 1-D vector.step, fully contiguous -> contiguous_chunk = 32 (the inner extent).
 // CHECK-LABEL: func.func @load_step_offsets(
 // CHECK: xegpu.load
-// CHECK-SAME: <{coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}>
+// CHECK-SAME: <{contiguous_chunk = 32 : i64}>
 // CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
 // CHECK-NOT: lane_data
 func.func @load_step_offsets(%ptr: i64) -> vector<32xf32> {
@@ -21,10 +21,10 @@ func.func @load_step_offsets(%ptr: i64) -> vector<32xf32> {
 }
 
 // -----
-// 2-D leading-1 dim: hint stamped on the load with factor = 2.
+// 2-D leading-1 dim: contiguity computed on the inner dim -> contiguous_chunk = 32.
 // CHECK-LABEL: func.func @load_2d_leading_unit(
 // CHECK: xegpu.load
-// CHECK-SAME: <{coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}>
+// CHECK-SAME: <{contiguous_chunk = 32 : i64}>
 // CHECK-SAME: : i64, vector<1x32xindex>, vector<1x32xi1> -> vector<1x32xf32>
 // CHECK-NOT: lane_data
 func.func @load_2d_leading_unit(%ptr: i64) -> vector<1x32xf32> {
@@ -37,12 +37,12 @@ func.func @load_2d_leading_unit(%ptr: i64) -> vector<1x32xf32> {
 }
 
 // -----
-// Stride-4 offsets: not coalescible, no hint stamped.
-// CHECK-LABEL: func.func @load_stride4_no_hint(
+// Stride-4 offsets: not contiguous, no attribute stamped.
+// CHECK-LABEL: func.func @load_stride4_no_chunk(
 // CHECK: xegpu.load
-// CHECK-NOT: coalesce_hint
+// CHECK-NOT: contiguous_chunk
 // CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
-func.func @load_stride4_no_hint(%ptr: i64) -> vector<32xf32> {
+func.func @load_stride4_no_chunk(%ptr: i64) -> vector<32xf32> {
   %c4 = arith.constant 4 : index
   %step = vector.step : vector<32xindex>
   %splat = vector.broadcast %c4 : index to vector<32xindex>
@@ -54,14 +54,13 @@ func.func @load_stride4_no_hint(%ptr: i64) -> vector<32xf32> {
 }
 
 // -----
-// All-equal offsets: classified as Broadcast by decide(); analysis returns
-// None at the stamping stage, so no hint is stamped (broadcast-load
-// rewrite was removed).
-// CHECK-LABEL: func.func @load_broadcast_offsets_no_hint(
+// All-equal offsets: inner dim is constant (not contiguous), so no attribute
+// is stamped.
+// CHECK-LABEL: func.func @load_broadcast_offsets_no_chunk(
 // CHECK: xegpu.load
-// CHECK-NOT: coalesce_hint
+// CHECK-NOT: contiguous_chunk
 // CHECK-SAME: : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
-func.func @load_broadcast_offsets_no_hint(%ptr: i64) -> vector<32xf32> {
+func.func @load_broadcast_offsets_no_chunk(%ptr: i64) -> vector<32xf32> {
   %offsets = arith.constant dense<0> : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
   %v = xegpu.load %ptr[%offsets], %mask
@@ -70,14 +69,13 @@ func.func @load_broadcast_offsets_no_hint(%ptr: i64) -> vector<32xf32> {
 }
 
 // -----
-// Store with vector.step offsets: hint stamped with factor = 2 on the store
-// op.
-// CHECK-LABEL: func.func @store_step_hint(
+// Store with vector.step offsets: contiguous_chunk = 32 stamped on the store.
+// CHECK-LABEL: func.func @store_step_chunk(
 // CHECK: xegpu.store
-// CHECK-SAME: <{coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}>
+// CHECK-SAME: <{contiguous_chunk = 32 : i64}>
 // CHECK-SAME: : vector<32xf32>, i64, vector<32xindex>, vector<32xi1>
 // CHECK-NOT: lane_data
-func.func @store_step_hint(%ptr: i64, %v: vector<32xf32>) {
+func.func @store_step_chunk(%ptr: i64, %v: vector<32xf32>) {
   %offsets = vector.step : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
   xegpu.store %v, %ptr[%offsets], %mask
@@ -86,15 +84,15 @@ func.func @store_step_hint(%ptr: i64, %v: vector<32xf32>) {
 }
 
 // -----
-// A user-provided coalesce_hint takes precedence: the analysis must not
-// overwrite it, even if its own heuristic would pick a different factor.
-// CHECK-LABEL: func.func @user_hint_preserved(
+// A pre-existing contiguous_chunk takes precedence: the analysis must not
+// overwrite it.
+// CHECK-LABEL: func.func @user_chunk_preserved(
 // CHECK: xegpu.load
-// CHECK-SAME: <{coalesce_hint = #xegpu.coalesce_hint<factor = 2 : i64>}>
-func.func @user_hint_preserved(%ptr: i64) -> vector<32xf32> {
+// CHECK-SAME: <{contiguous_chunk = 2 : i64}>
+func.func @user_chunk_preserved(%ptr: i64) -> vector<32xf32> {
   %offsets = vector.step : vector<32xindex>
   %mask = arith.constant dense<true> : vector<32xi1>
-  %v = xegpu.load %ptr[%offsets], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 2>}>
+  %v = xegpu.load %ptr[%offsets], %mask <{contiguous_chunk = 2 : i64}>
       : i64, vector<32xindex>, vector<32xi1> -> vector<32xf32>
   return %v : vector<32xf32>
 }
diff --git a/mlir/test/Dialect/XeGPU/invalid.mlir b/mlir/test/Dialect/XeGPU/invalid.mlir
index 6e7908836043f..c197788deeb38 100644
--- a/mlir/test/Dialect/XeGPU/invalid.mlir
+++ b/mlir/test/Dialect/XeGPU/invalid.mlir
@@ -774,25 +774,17 @@ func.func @dpas_mx_scale_b_layout_not_distributable(%a : vector<8x16xf8E5M2>, %b
 }
 
 // -----
-func.func @coalesce_hint_not_power_of_two(%src: i64, %offset: vector<16xindex>, %mask: vector<16xi1>) {
-  // expected-error at +1 {{'factor' : 6 must be a power of two}}
-  %val = xegpu.load %src[%offset], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 6>}>
+func.func @contiguous_chunk_too_small(%src: i64, %offset: vector<16xindex>, %mask: vector<16xi1>) {
+  // expected-error at +1 {{contiguous_chunk 1 must be >= 2}}
+  %val = xegpu.load %src[%offset], %mask <{contiguous_chunk = 1 : i64}>
       : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
   return
 }
 
 // -----
-func.func @coalesce_hint_factor_too_small(%src: i64, %offset: vector<16xindex>, %mask: vector<16xi1>) {
-  // expected-error at +1 {{'factor' : 1 must be >= 2}}
-  %val = xegpu.load %src[%offset], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 1>}>
-      : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
-  return
-}
-
-// -----
-func.func @coalesce_hint_factor_does_not_divide(%src: i64, %offset: vector<6xindex>, %mask: vector<6xi1>) {
-  // expected-error at +1 {{coalesce_hint factor 4 must divide the innermost offsets dim 6}}
-  %val = xegpu.load %src[%offset], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 4>}>
+func.func @contiguous_chunk_exceeds_inner(%src: i64, %offset: vector<6xindex>, %mask: vector<6xi1>) {
+  // expected-error at +1 {{contiguous_chunk 8 must not exceed the innermost offsets dim 6}}
+  %val = xegpu.load %src[%offset], %mask <{contiguous_chunk = 8 : i64}>
       : i64, vector<6xindex>, vector<6xi1> -> vector<6xf32>
   return
 }
diff --git a/mlir/test/Dialect/XeGPU/ops.mlir b/mlir/test/Dialect/XeGPU/ops.mlir
index d0ba3c2119072..7adbbe7dea2ba 100644
--- a/mlir/test/Dialect/XeGPU/ops.mlir
+++ b/mlir/test/Dialect/XeGPU/ops.mlir
@@ -452,19 +452,19 @@ gpu.func @subgroup_store_offset_1(%dest: memref<?xf16>) {
   gpu.return
 }
 
-// CHECK: gpu.func @load_coalesce_hint(%[[arg0:.*]]: i64, %[[arg1:.*]]: vector<16xindex>, %[[arg2:.*]]: vector<16xi1>) {
-gpu.func @load_coalesce_hint(%src: i64, %offset: vector<16xindex>, %mask: vector<16xi1>) {
-  // A user-provided `coalesce_hint` round-trips through the optional op attribute.
-  // CHECK: xegpu.load %[[arg0]][%[[arg1]]], %[[arg2]] <{coalesce_hint = #xegpu.coalesce_hint<factor = 4 : i64>}> : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
-  %val = xegpu.load %src[%offset], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 4>}>
+// CHECK: gpu.func @load_contiguous_chunk(%[[arg0:.*]]: i64, %[[arg1:.*]]: vector<16xindex>, %[[arg2:.*]]: vector<16xi1>) {
+gpu.func @load_contiguous_chunk(%src: i64, %offset: vector<16xindex>, %mask: vector<16xi1>) {
+  // A user-provided `contiguous_chunk` round-trips through the optional op attribute.
+  // CHECK: xegpu.load %[[arg0]][%[[arg1]]], %[[arg2]] <{contiguous_chunk = 4 : i64}> : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
+  %val = xegpu.load %src[%offset], %mask <{contiguous_chunk = 4 : i64}>
       : i64, vector<16xindex>, vector<16xi1> -> vector<16xf32>
   gpu.return
 }
 
-// CHECK: gpu.func @store_coalesce_hint(%[[arg0:.*]]: vector<16xf32>, %[[arg1:.*]]: i64, %[[arg2:.*]]: vector<16xindex>, %[[arg3:.*]]: vector<16xi1>) {
-gpu.func @store_coalesce_hint(%val: vector<16xf32>, %dest: i64, %offset: vector<16xindex>, %mask: vector<16xi1>) {
-  // CHECK: xegpu.store %[[arg0]], %[[arg1]][%[[arg2]]], %[[arg3]] <{coalesce_hint = #xegpu.coalesce_hint<factor = 4 : i64>}> : vector<16xf32>, i64, vector<16xindex>, vector<16xi1>
-  xegpu.store %val, %dest[%offset], %mask <{coalesce_hint = #xegpu.coalesce_hint<factor = 4>}>
+// CHECK: gpu.func @store_contiguous_chunk(%[[arg0:.*]]: vector<16xf32>, %[[arg1:.*]]: i64, %[[arg2:.*]]: vector<16xindex>, %[[arg3:.*]]: vector<16xi1>) {
+gpu.func @store_contiguous_chunk(%val: vector<16xf32>, %dest: i64, %offset: vector<16xindex>, %mask: vector<16xi1>) {
+  // CHECK: xegpu.store %[[arg0]], %[[arg1]][%[[arg2]]], %[[arg3]] <{contiguous_chunk = 4 : i64}> : vector<16xf32>, i64, vector<16xindex>, vector<16xi1>
+  xegpu.store %val, %dest[%offset], %mask <{contiguous_chunk = 4 : i64}>
       : vector<16xf32>, i64, vector<16xindex>, vector<16xi1>
   gpu.return
 }
diff --git a/mlir/test/lib/Dialect/XeGPU/TestXeGPUTransforms.cpp b/mlir/test/lib/Dialect/XeGPU/TestXeGPUTransforms.cpp
index f24b4dccaab04..f7b768db36460 100644
--- a/mlir/test/lib/Dialect/XeGPU/TestXeGPUTransforms.cpp
+++ b/mlir/test/lib/Dialect/XeGPU/TestXeGPUTransforms.cpp
@@ -16,12 +16,16 @@
 #include "mlir/Dialect/XeGPU/Transforms/Transforms.h"
 #include "mlir/Dialect/XeGPU/Transforms/XeGPULayoutImpl.h"
 #include "mlir/Dialect/XeGPU/Utils/XeGPUUtils.h"
+#include "mlir/Dialect/XeGPU/uArch/IntelGpuXe2.h"
 #include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/Value.h"
 #include "mlir/Pass/Pass.h"
 #include "mlir/Pass/PassManager.h"
 #include "mlir/Transforms/DialectConversion.h"
 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+#include "llvm/ADT/ScopeExit.h"
+#include "llvm/ADT/bit.h"
+#include "llvm/Support/MathExtras.h"
 #include "llvm/Support/raw_ostream.h"
 #include <optional>
 
@@ -455,7 +459,7 @@ struct TestXeGPUCoalesceGatherScatter
   }
 
   StringRef getDescription() const final {
-    return "Test the XeGPU coalesce-gather-scatter analysis + apply APIs.";
+    return "Test the XeGPU contiguity analysis and its coalescing consumer.";
   }
 
   void getDependentDialects(::mlir::DialectRegistry &registry) const override {
@@ -475,16 +479,163 @@ struct TestXeGPUCoalesceGatherScatter
 
   Option<bool> analyzeOnly{
       *this, "analyze-only",
-      llvm::cl::desc("Only run the analysis (stamp xegpu.coalesce_hint "
+      llvm::cl::desc("Only run the analysis (stamp contiguous_chunk "
                      "attributes); do not apply."),
       llvm::cl::init(false)};
 
   void runOnOperation() override {
-    xegpu::CoalesceGatherScatterAnalysisOptions options;
-    options.maxChunkSize = maxChunkSize;
-    xegpu::runCoalesceGatherScatterAnalysis(getOperation(), options);
-    if (!analyzeOnly)
-      xegpu::coalesceGatherScatter(getOperation());
+    xegpu::runCoalesceGatherScatterAnalysis(getOperation());
+    if (analyzeOnly)
+      return;
+    getOperation()->walk([&](Operation *op) {
+      if (auto load = dyn_cast<xegpu::LoadGatherOp>(op))
+        applyContiguousChunk(load, maxChunkSize);
+      else if (auto store = dyn_cast<xegpu::StoreScatterOp>(op))
+        applyContiguousChunk(store, maxChunkSize);
+    });
+  }
+
+private:
+  /// Largest power-of-two `<= bound` that divides `numLanes`.
+  static int64_t largestPow2Divisor(int64_t numLanes, int64_t bound) {
+    if (bound < 2 || numLanes < 2)
+      return 1;
+    int64_t f = std::min<int64_t>(bound, numLanes);
+    // Round down to power of 2.
+    if (!llvm::isPowerOf2_64(f))
+      f = static_cast<int64_t>(llvm::bit_floor(static_cast<uint64_t>(f)));
+    while (f >= 2) {
+      if (numLanes % f == 0)
+        return f;
+      f /= 2;
+    }
+    return 1;
+  }
+
+  /// Look up the subgroup size from the enclosing gpu.module's xevm.target.
+  /// Falls back to 16 when no target chip is found or the chip is unknown,
+  /// matching the typical Intel Xe2 default.
+  static unsigned lookupSubgroupSize(Operation *op) {
+    const auto *uArch =
+        xegpu::uArch::getUArch(xegpu::getChipStr(op).value_or(""));
+    return uArch ? static_cast<unsigned>(uArch->getSubgroupSize()) : 16u;
+  }
+
+  /// Build a `lane_layout`/`lane_data`/`inst_data` layout of rank `rank`, with
+  /// the given lane_layout / lane_data on the innermost dim (1 elsewhere).
+  /// `inst_data` is `lane_layout * lane_data` per dim.
+  static xegpu::LayoutAttr buildLaneDataLayout(MLIRContext *ctx, unsigned rank,
+                                               int64_t innerLaneLayout,
+                                               int64_t innerLaneData) {
+    SmallVector<int32_t> laneLayout(rank, 1);
+    SmallVector<int32_t> laneData(rank, 1);
+    SmallVector<int32_t> instData(rank, 1);
+    laneLayout.back() = static_cast<int32_t>(innerLaneLayout);
+    laneData.back() = static_cast<int32_t>(innerLaneData);
+    instData.back() = static_cast<int32_t>(innerLaneLayout * innerLaneData);
+    return xegpu::LayoutAttr::get(ctx, instData, laneLayout, laneData);
+  }
+
+  /// Returns true if `mask` is a constant `dense<true>` vector.
+  static bool isAllTrueMask(Value mask) {
+    auto vecTy = dyn_cast<VectorType>(mask.getType());
+    if (!vecTy)
+      return false;
+    auto cst = mask.getDefiningOp<arith::ConstantOp>();
+    if (!cst)
+      return false;
+    auto dense = dyn_cast<DenseIntElementsAttr>(cst.getValue());
+    if (!dense || !dense.isSplat())
+      return false;
+    return dense.getSplatValue<APInt>().getBoolValue();
+  }
+
+  /// True when `op` (a gather load or scatter store) is tied to a
+  /// `vector.multi_reduction`: a load whose result feeds a reduction, or a
+  /// store whose stored value comes from one (through layout-neutral /
+  /// elementwise / insert glue). Coalescing such an access is gated off here;
+  /// it requires reduction-aware layout handling added in follow-up PRs.
+  template <typename OpTy>
+  static bool isReductionTied(OpTy op) {
+    if constexpr (std::is_same_v<OpTy, xegpu::StoreScatterOp>) {
+      SmallVector<Value, 8> worklist{op.getValue()};
+      llvm::SmallPtrSet<Operation *, 16> seen;
+      unsigned steps = 0;
+      while (!worklist.empty() && steps++ < 64) {
+        Value v = worklist.pop_back_val();
+        Operation *def = v.getDefiningOp();
+        if (!def || !seen.insert(def).second)
+          continue;
+        if (isa<vector::MultiDimReductionOp>(def))
+          return true;
+        if (isa<vector::ShapeCastOp, vector::BitCastOp, xegpu::ConvertLayoutOp,
+                vector::InsertOp, vector::InsertStridedSliceOp>(def) ||
+            OpTrait::hasElementwiseMappableTraits(def))
+          for (Value operand : def->getOperands())
+            if (isa<VectorType>(operand.getType()))
+              worklist.push_back(operand);
+      }
+      return false;
+    } else {
+      for (Operation *user : op->getUsers()) {
+        Operation *u = user;
+        while (
+            u &&
+            isa<vector::ShapeCastOp, vector::BitCastOp, xegpu::ConvertLayoutOp>(
+                u)) {
+          if (u->getNumResults() != 1 || u->getResult(0).use_empty())
+            break;
+          u = *u->getResult(0).getUsers().begin();
+        }
+        if (u && isa<vector::MultiDimReductionOp>(u))
+          return true;
+      }
+      return false;
+    }
+  }
+
+  /// Turn a stamped `contiguous_chunk` on `op` into a lane_layout / lane_data /
+  /// inst_data layout, capped by `maxChunkSize`, then remove the attribute.
+  /// Skips ops with a non-uniform mask, an existing lane_data, or a reduction
+  /// tie — these are coalescing concerns, not properties of the offsets.
+  template <typename OpTy>
+  static void applyContiguousChunk(OpTy op, unsigned maxChunkSize) {
+    std::optional<uint64_t> chunk = op.getContiguousChunk();
+    if (!chunk)
+      return;
+    auto cleanup = llvm::scope_exit([&] { op.removeContiguousChunkAttr(); });
+
+    auto offsetsTy = dyn_cast<VectorType>(op.getOffsets().getType());
+    auto valueTy = op.getValueType();
+    if (!offsetsTy || !valueTy || offsetsTy.getNumElements() <= 1)
+      return;
+    if (!isAllTrueMask(op.getMask()))
+      return;
+    if (auto layout = op.getLayoutAttr())
+      if (!layout.getEffectiveLaneDataAsInt().empty())
+        return;
+    if (isReductionTied(op))
+      return;
+
+    int64_t inner = offsetsTy.getShape().back();
+    unsigned subgroupSize = lookupSubgroupSize(op);
+    // lane_layout default: min(subgroupSize, inner) rounded to a divisor.
+    int64_t laneLayout =
+        largestPow2Divisor(inner, std::min<int64_t>(subgroupSize, inner));
+    if (laneLayout < 1)
+      return;
+    int64_t perLane = inner / laneLayout;
+    // lane_data = min(contiguity, maxChunkSize, perLane), pow2 divisor of
+    // perLane.
+    int64_t bound =
+        std::min<int64_t>({static_cast<int64_t>(*chunk),
+                           static_cast<int64_t>(maxChunkSize), perLane});
+    int64_t factor = largestPow2Divisor(perLane, bound);
+    if (factor < 2)
+      return;
+
+    op.setLayoutAttr(buildLaneDataLayout(op.getContext(), valueTy.getRank(),
+                                         laneLayout, factor));
   }
 };
 

>From b54ebdad62616e3d7ed33173e5321950badaf22c Mon Sep 17 00:00:00 2001
From: "Shahneous Bari, Md Abdullah" <md.abdullah.shahneous.bari at intel.com>
Date: Wed, 24 Jun 2026 15:41:43 +0000
Subject: [PATCH 7/7] Fix build/format after rename to XeGPUContiguityAnalysis.

- CMakeLists.txt referenced the old XeGPUCoalesceGatherScatter.cpp name,
  breaking the build (CI failure).
- clang-format VectorToXeGPU.cpp (reflow after the longer
  contiguous_chunk builder label).
- Update the renamed file's banner comment.

Co-Authored-By: Claude Opus 4.8 <noreply at anthropic.com>
---
 .../lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp | 14 +++++++-------
 mlir/lib/Dialect/XeGPU/Transforms/CMakeLists.txt   |  2 +-
 .../XeGPU/Transforms/XeGPUContiguityAnalysis.cpp   |  2 +-
 3 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp b/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp
index a428a9d82e618..bcb2b3530e2f7 100644
--- a/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp
+++ b/mlir/lib/Conversion/VectorToXeGPU/VectorToXeGPU.cpp
@@ -527,13 +527,13 @@ static LogicalResult lowerToScatteredStoreOp(vector::TransferWriteOp writeOp,
   Value mask = vector::ConstantMaskOp::create(
       rewriter, loc, VectorType::get(vectorShape, rewriter.getI1Type()),
       vectorShape);
-  xegpu::StoreScatterOp::create(rewriter, loc, writeOp.getVector(), flatMemref,
-                                localOffsets, mask,
-                                /*chunk_size=*/IntegerAttr{},
-                                /*l1_hint=*/xegpu::CachePolicyAttr{},
-                                /*l2_hint=*/xegpu::CachePolicyAttr{},
-                                /*l3_hint=*/xegpu::CachePolicyAttr{},
-                                /*layout=*/nullptr, /*contiguous_chunk=*/nullptr);
+  xegpu::StoreScatterOp::create(
+      rewriter, loc, writeOp.getVector(), flatMemref, localOffsets, mask,
+      /*chunk_size=*/IntegerAttr{},
+      /*l1_hint=*/xegpu::CachePolicyAttr{},
+      /*l2_hint=*/xegpu::CachePolicyAttr{},
+      /*l3_hint=*/xegpu::CachePolicyAttr{},
+      /*layout=*/nullptr, /*contiguous_chunk=*/nullptr);
   rewriter.eraseOp(writeOp);
   return success();
 }
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/CMakeLists.txt b/mlir/lib/Dialect/XeGPU/Transforms/CMakeLists.txt
index 1b013400cff34..2ed81ae05ab34 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/CMakeLists.txt
+++ b/mlir/lib/Dialect/XeGPU/Transforms/CMakeLists.txt
@@ -1,7 +1,7 @@
 add_mlir_dialect_library(MLIRXeGPUTransforms
   XeGPUArrayLengthOptimization.cpp
   XeGPUBlocking.cpp
-  XeGPUCoalesceGatherScatter.cpp
+  XeGPUContiguityAnalysis.cpp
   XeGPUSgToLaneDistribute.cpp
   XeGPUUnroll.cpp
   XeGPUWgToSgDistribute.cpp
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUContiguityAnalysis.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUContiguityAnalysis.cpp
index 1c67503f7445a..5efbed59f38ee 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUContiguityAnalysis.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUContiguityAnalysis.cpp
@@ -1,4 +1,4 @@
-//===- XeGPUCoalesceGatherScatter.cpp - Coalesce scatter accesses --------===//
+//===- XeGPUContiguityAnalysis.cpp - Offset contiguity analysis ---------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.



More information about the Mlir-commits mailing list