[Path B] Restore baseline 1-shim-per-compute-col placement

CI Triton 8x4 routing failure root cause: Path B's centroid-driven shim
placement put 6 shim tiles clustered at cols 0-5, leaving compute cols
6-7 with no nearby shim. mlir-aie's pathfinder then can't find a legal
route through the network. Baseline (pre-Xilinx#1605, pre-Path-B) deterministically
produced 8 shim cols (one per active compute col) via the same-column
heuristic, which routed cleanly.

Fix has three pieces; this commit lands the AIR side and bumps the
mlir-aie pin to pull in the third:

1. AIR (this commit): emit shim LTOs as `aie.logical_tile<ShimNOCTile>(
   compute_col, ?)` whenever the device has a ShimNOC tile at that col.
   On AIE1 (sparse ShimNOC at cols 2/6/10) the hint stays unset and the
   placer falls back to centroid placement, preserving existing behavior.

2. AIR (this commit): scope LTO grouping to same-col candidates. Without
   this, the first shim allocation creates an LTO and all subsequent
   allocations reuse it regardless of compute col, so the per-col hint is
   never honored. Now allocations only group onto an LTO whose col hint
   matches their compute col.

3. mlir-aie #3064 (already merged at 45915e4): extend
   `findTileWithCapacity` from sweep-right-only to bidirectional sweep.
   Bumps utils/clone-mlir-aie.sh from b37dc33 to 45915e4 to pick this up.

Verified locally: Path B now produces bit-identical placement to
baseline trunk for the failing Triton 8x4 workload — 48 unique tiles,
8 shim cols at 0-7, 8 memtile cols at 0-7, 32 compute cores at rows
2-5. Lit suite: 370/372 pass (only 2 pre-existing AIRToROCDL failures
unrelated to Path B).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: erwei-xilinx

mlir/lib/Conversion/AIRToAIESchedulingUtils.cpp

@@ -1068,13 +1068,27 @@ air::ShimDMAAllocator::allocNewDmaChannel(air::MemcpyInterface &memcpyOp,
         return c;
     return -1;
   };
+  // Only reuse an existing LTO if its col hint matches `col` (the
+  // compute-side column). This preserves baseline's "1 shim per active
+  // compute col" placement under the LTO model: each compute col gets
+  // its own shim LTO (with `(col, ?)` hint), so the placer + bidirectional
+  // sweep (mlir-aie #3064) can spread shims under each compute col rather
+  // than clustering near the centroid.
   for (auto *side : {&mm2s_allocs, &s2mm_allocs}) {
     for (auto &t : *side) {
       auto cand = dyn_cast<AIE::LogicalTileOp>(t.dma_tile.getOperation());
       if (!cand)
         continue;
       if (cand.getTileType() != AIE::AIETileType::ShimNOCTile)
         continue;
+      auto candCol = cand.getCol();
+      if (col >= 0) {
+        if (!candCol || (int)*candCol != col)
+          continue;
+      } else {
+        if (candCol)
+          continue;
+      }
       int c = pickChannelForLTO(cand);
       if (c < 0)
         continue;
@@ -1094,9 +1108,14 @@ air::ShimDMAAllocator::allocNewDmaChannel(air::MemcpyInterface &memcpyOp,
       else
         break;
     }
+    auto *ctx = b.getContext();
+    const auto &tm = device.getTargetModel();
+    IntegerAttr colAttr =
+        (col >= 0 && col < tm.columns() && tm.isShimNOCTile(col, 0))
+            ? IntegerAttr::get(IntegerType::get(ctx, 32), col)
+            : IntegerAttr();
     tileLT = AIE::LogicalTileOp::create(b, device.getLoc(),
-                                        AIE::AIETileType::ShimNOCTile,
-                                        /*col=*/IntegerAttr(),
+                                        AIE::AIETileType::ShimNOCTile, colAttr,
                                         /*row=*/IntegerAttr(),
                                         /*allocation_scheme=*/StringAttr());
     dma_channel = 0;

utils/clone-mlir-aie.sh

@@ -14,8 +14,8 @@
 #
 ##===----------------------------------------------------------------------===##
 
-export HASH=b37dc33d41511684fd4eef1b8ac2e3f74fd5f169
-DATETIME=2026050821
+export HASH=45915e410804c1859f7fffa3a3369485970577e8
+DATETIME=2026051117
 WHEEL_VERSION=0.0.1.$DATETIME+${HASH:0:7}
 
 if [ x"$1" == x--get-wheel-version ]; then