a.diff

diff --git a/compiler/src/iree/compiler/Codegen/LLVMCPU/KernelDispatch.cpp b/compiler/src/iree/compiler/Codegen/LLVMCPU/KernelDispatch.cpp
index 002c34f425..e4dd03268f 100644
--- a/compiler/src/iree/compiler/Codegen/LLVMCPU/KernelDispatch.cpp
+++ b/compiler/src/iree/compiler/Codegen/LLVMCPU/KernelDispatch.cpp
@@ -1506,7 +1506,8 @@ static void setX86VectorTileSizes(linalg::GenericOp genericOp,
                                   SmallVectorImpl<int64_t> &vecTileSizes) {
   vecTileSizes.append(numLoops, 0);
   SmallVector<int64_t> staticLoopRanges = genericOp.getStaticLoopRanges();
-  for (auto loopNum : llvm::seq<unsigned>(0, numLoops)) {
+  int64_t maxRemainingVecTileSize = 1024;
+  for (auto loopNum : llvm::reverse(llvm::seq<unsigned>(0, numLoops))) {
     if (distTileSizes[loopNum]) {
       vecTileSizes[loopNum] = getMaxVectorTileSize(
           distTileSizes[loopNum], minTileSizes[loopNum], minTileSizes[loopNum],
@@ -1518,7 +1519,15 @@ static void setX86VectorTileSizes(linalg::GenericOp genericOp,
       vecTileSizes[loopNum] =
           staticLoopRanges[loopNum] == 1 ? 0 : minTileSizes[loopNum];
     }
+    if (vecTileSizes[loopNum] > maxRemainingVecTileSize) {
+      vecTileSizes[loopNum] = maxRemainingVecTileSize;
+      maxRemainingVecTileSize = 1;
+    } else {
+      maxRemainingVecTileSize = std::max<int64_t>(1, maxRemainingVecTileSize / vecTileSizes[loopNum]);
+    }
   }
+  llvm::dbgs() << "setX86VectorTileSizes:\n";
+  for (auto x : vecTileSizes) llvm::dbgs() << "    " << x << "\n";  
 }