bump llvm

Author: akroviakov

include/gc/Dialect/LLVMIR/XeVMOps.td

@@ -70,7 +70,7 @@ def XeVM_L1StoreCacheControl : XeVM_StoreCacheControl<"L1">;
 def XeVM_L3StoreCacheControl : XeVM_StoreCacheControl<"L3">;
 
 def XeVM_BlockLoad2dOp : XeVM_Op<"blockload2d">,
-  Results<(outs FixedVectorOf<[XeVM_ElemType]>:$res)>,
+  Results<(outs FixedVectorOfRankAndType<[1,2,3], [XeVM_ElemType]>:$res)>,
   Arguments<(ins
     Arg<LLVM_AnyPointer, "", [MemRead]>:$ptr,
     I32:$base_width,
@@ -137,7 +137,7 @@ def XeVM_BlockStore2dOp : XeVM_Op<"blockstore2d">,
     I32Attr:$tile_width,
     I32Attr:$tile_height,
     I32Attr:$v_blocks,
-    FixedVectorOf<[XeVM_ElemType]>:$stored_val,
+    FixedVectorOfRankAndType<[1, 2, 3], [XeVM_ElemType]>:$stored_val,
     DefaultValuedAttr<XeVM_L1StoreCacheControl, "::mlir::xevm::L1StoreCacheControl::DEFAULT">:$l1_cache_control,
     DefaultValuedAttr<XeVM_L3StoreCacheControl, "::mlir::xevm::L3StoreCacheControl::DEFAULT">:$l3_cache_control
   )> {
@@ -243,7 +243,7 @@ def XeVM_PrecisionTypeAttr : I32EnumAttr<"PrecisionType",
 }
 
 def XeVM_DPASOp : XeVM_Op<"dpas">,
-  Results<(outs FixedVectorOf<[XeVM_MatrixElemType]>:$d)>,
+  Results<(outs FixedVectorOfRankAndType<[1], [XeVM_MatrixElemType]>:$d)>,
   Arguments<(ins
     FixedVectorOfRankAndType<[1], [XeVM_MatrixElemType]>:$c,
     FixedVectorOfRankAndType<[1], [XeVM_MatrixElemType]>:$a,

include/gc/Transforms/Microkernel/BrgemmRuntimeUtils.h

@@ -27,13 +27,13 @@ static inline int64_t getDnnlDataTypeVal(RewriterBase &rewriter,
   auto context = rewriter.getContext();
   auto tattr = dyn_cast_or_null<TypeAttr>(attr);
   assert(tattr);
-  if (tattr == TypeAttr::get(FloatType::getF32(context))) {
+  if (tattr == TypeAttr::get(Float32Type::get(context))) {
     return static_cast<int64_t>(dnnl_f32);
-  } else if (tattr == TypeAttr::get(FloatType::getF64(context))) {
+  } else if (tattr == TypeAttr::get(Float64Type::get(context))) {
     return static_cast<int64_t>(dnnl_f64);
-  } else if (tattr == TypeAttr::get(FloatType::getBF16(context))) {
+  } else if (tattr == TypeAttr::get(BFloat16Type::get(context))) {
     return static_cast<int64_t>(dnnl_bf16);
-  } else if (tattr == TypeAttr::get(FloatType::getF16(context))) {
+  } else if (tattr == TypeAttr::get(Float16Type::get(context))) {
     return static_cast<int64_t>(dnnl_f16);
   } else if (tattr == TypeAttr::get(
                           IntegerType::get(context, 32, IntegerType::Signed))) {

include/gc/Transforms/Utils/StructuredOpMatcher.h

@@ -163,7 +163,7 @@ struct HasStaticStrides {
     SmallVector<int64_t> strides;
     if (auto memRefType = dyn_cast_or_null<MemRefType>(operandType)) {
       int64_t offset;
-      if (failed(getStridesAndOffset(memRefType, strides, offset)))
+      if (failed(memRefType.getStridesAndOffset(strides, offset)))
         return false;
       if (llvm::any_of(strides, [](int64_t stride) {
             return stride == ShapedType::kDynamic;
@@ -244,7 +244,8 @@ struct NumDpsInits {
 // Callable object to validate number of input operands for `op`.
 struct NumDpsInputs {
   NumDpsInputs() = delete;
-  explicit NumDpsInputs(std::function<bool(size_t)> fun) : fun(std::move(fun)){};
+  explicit NumDpsInputs(std::function<bool(size_t)> fun)
+      : fun(std::move(fun)){};
 
   bool operator()(Operation *op) {
     if (auto linalgOp = dyn_cast_or_null<linalg::LinalgOp>(op))

lib/gc/Dialect/Linalgx/Utils.cpp

@@ -385,7 +385,7 @@ bool isGenericAttrEquivalent(linalg::GenericOp op, ShapedType shapeA,
     DenseMap<AffineExpr, AffineExpr> replaceMap;
     std::map<unsigned, utils::IteratorType> iterMap;
     // get shape-to-loop map
-    AffineMap inverse = inversePermutation(concatAffineMaps(inMaps));
+    AffineMap inverse = inversePermutation(concatAffineMaps(inMaps, context));
     assert(inverse && "shape-to-loops map to be non-null");
     assert(dimSize == inverse.getResults().size());
     // renumber the dim id based on shape-to-loop map

lib/gc/Dialect/Microkernel/MicrokernelOps.cpp

@@ -551,11 +551,11 @@ static LogicalResult verifyBrgemmDataTypes(ArrayAttr dtypes,
 
   auto context = op.getContext();
 
-#define FTAttr(t) TypeAttr::get(FloatType::get##t(context))
+#define FTAttr(t) TypeAttr::get(t::get(context))
 #define ITAttr(s, w) TypeAttr::get(IntegerType::get(context, w, IntegerType::s))
   SmallVector<std::pair<TypeAttr, TypeAttr>> validDataTypes = {
-      {FTAttr(F32), FTAttr(F32)},
-      {FTAttr(BF16), FTAttr(BF16)},
+      {FTAttr(Float32Type), FTAttr(Float32Type)},
+      {FTAttr(BFloat16Type), FTAttr(BFloat16Type)},
       {ITAttr(Unsigned, 8), ITAttr(Signed, 8)},
       {ITAttr(Signed, 8), ITAttr(Unsigned, 8)},
       {ITAttr(Unsigned, 8), ITAttr(Unsigned, 8)},

lib/gc/ExecutionEngine/GPURuntime/ocl/GpuOclRuntime.cpp

@@ -718,7 +718,7 @@ StringRef createStaticMain(OpBuilder &builder, ModuleOp &module,
       auto offsetPtr = constArgs.end();
       constArgs.emplace_back(0);
       constArgs.append(shape.begin(), shape.end());
-      if (failed(getStridesAndOffset(type, constArgs, *offsetPtr))) {
+      if (failed(type.getStridesAndOffset(constArgs, *offsetPtr))) {
         gcLogD("Failed to get strides and offset of arg", i,
                " of the function ", funcName.begin());
         return {};
@@ -929,8 +929,9 @@ OclModuleBuilder::build(const OclRuntime::Ext &ext) {
             builder.getI64IntegerAttr(static_cast<int64_t>(wgSize)));
     TargetDeviceSpecInterface devSpec =
         TargetDeviceSpecAttr::get(ctx, dltiAttrs);
-    auto sysSpec =
-        TargetSystemSpecAttr::get(ctx, ArrayRef(std::pair(devStr, devSpec)));
+    DataLayoutEntryInterface dl =
+        DataLayoutEntryAttr::get(ctx, devStr, devSpec);
+    auto sysSpec = TargetSystemSpecAttr::get(ctx, ArrayRef(dl));
     mod = mlirModule.clone();
     mod.getOperation()->setAttr("#dlti.sys_spec", sysSpec);
     PassManager pm{ctx};

lib/gc/Transforms/DeepTileContractionOp.cpp

@@ -405,7 +405,7 @@ generateOuterLoop(RewriterBase &b, linalg::LinalgOp linalgOp,
         // the extra copy generated by bufferization. So remove the dummy loop
         // at this early stage.
         if (!isDummyLoop(tilingResult->loops.back())) {
-          b.replaceOp(currentOp, tilingResult->replacements);
+          b.replaceOp(currentOp, tilingResult->mergeResult.replacements);
           currentOp = dyn_cast<linalg::LinalgOp>(tilingResult->tiledOps.back());
           if (iteratorTypes[d] == mlir::utils::IteratorType::reduction)
             result.reductionLoops.push_back(tilingResult->loops.back());
@@ -477,7 +477,7 @@ generateOuterLoop(RewriterBase &b, linalg::LinalgOp linalgOp,
             b, cast<TilingInterface>(currentOp.getOperation()), tileOption);
         if (failed(tilingResult))
           return failure();
-        b.replaceOp(currentOp, tilingResult->replacements);
+        b.replaceOp(currentOp, tilingResult->mergeResult.replacements);
         currentOp = dyn_cast<linalg::LinalgOp>(tilingResult->tiledOps.back());
       }
     }

lib/gc/Transforms/IterativeTilingAndFusion.cpp

@@ -813,7 +813,7 @@ void iterativeTilingAndFusionUntilExhaustion(
               defaultTilingOfType(rewriter, op, isaOpTy, cfg);
           if (succeeded(tilingResult)) {
             tiledOps.insert(tilingResult->tiledOps[0]);
-            rewriter.replaceOp(op, tilingResult->replacements);
+            rewriter.replaceOp(op, tilingResult->mergeResult.replacements);
             break;
           }
         }

lib/gc/Transforms/MemRefToCPURuntime.cpp

@@ -51,7 +51,7 @@ uint64_t getMemRefSizeInBytes(MemRefType memrefType) {
   if (!layout.isIdentity()) {
     int64_t offset;
     SmallVector<int64_t, 4> strides;
-    if (failed(getStridesAndOffset(memrefType, strides, offset))) {
+    if (failed(memrefType.getStridesAndOffset(strides, offset))) {
       return UINT64_MAX;
     }
 

lib/gc/Transforms/Microkernel/ConvertMicrokernelToDnnlFunc.cpp

@@ -63,7 +63,7 @@ class ConvertBrgemmDispatchOpRewriter
     SmallVector<Value, 10> operands;
     SmallVector<Type, 10> operandTypes;
     IntegerType integer64 = IntegerType::get(rewriter.getContext(), 64);
-    FloatType float32 = FloatType::getF32(rewriter.getContext());
+    FloatType float32 = Float32Type::get(rewriter.getContext());
 
     // M, N, K, LDA, LDB, LDC, stride_a, stride_b
     // they are in the same order with BrgemmDispatchOp inputs