Update llvm sha

.github/workflows/build.yml

@@ -82,11 +82,11 @@ jobs:
         id: cache-llvm-mlir
         uses: actions/cache@v3
         env:
-          CACHE_NUMBER: 1  # Increase to reset cache
+          LLVM_CACHE_NUMBER: 1  # Increase to reset cache
         with:
           path: |
             /home/runner/work/llvm-mlir/_mlir_install/**
-          key: ${{ runner.os }}-build-llvm-${{ env.CACHE_NUMBER }}-${{ env.LLVM_SHA }}
+          key: ${{ runner.os }}-build-llvm-${{ env.LLVM_CACHE_NUMBER }}-${{ env.LLVM_SHA }}
 
       - name: Download TBB
         if: steps.cache-tbb.outputs.cache-hit != 'true'
@@ -158,6 +158,7 @@ jobs:
             -DLLVM_ENABLE_RTTI=ON                                          \
             -DLLVM_USE_LINKER=gold                                         \
             -DLLVM_INSTALL_UTILS=ON                                        \
+            -DLLVM_ENABLE_ZSTD=OFF                                         \
             -DCMAKE_INSTALL_PREFIX=/home/runner/work/llvm-mlir/_mlir_install || exit 1
           cmake --build . -j ${np} || exit 1
           cmake --install . || exit 1

llvm-sha.txt

@@ -1 +1 @@
-fbb51ac0ba1007250cba384d930ae965e395b657
+78650b78618840563a05d840794519422998adbb

mlir/lib/Conversion/SCFToAffine/SCFToAffine.cpp

@@ -61,7 +61,7 @@ class SCFParallelLowering : public OpRewritePattern<scf::ParallelOp> {
         llvm::makeArrayRef(newSteps));
 
     // Steal the body of the old affine for op.
-    newPloop.region().takeBody(op.getRegion());
+    newPloop.getRegion().takeBody(op.getRegion());
 
     Operation *yieldOp = newPloop.getBody()->getTerminator();
     assert(yieldOp);

mlir/lib/Dialect/plier_util/dialect.cpp

@@ -272,48 +272,6 @@ struct GenGlobalId : public mlir::OpRewritePattern<mlir::arith::AddIOp> {
   }
 };
 
-struct InvertCmpi : public mlir::OpRewritePattern<mlir::arith::CmpIOp> {
-  using OpRewritePattern::OpRewritePattern;
-
-  mlir::LogicalResult
-  matchAndRewrite(mlir::arith::CmpIOp op,
-                  mlir::PatternRewriter &rewriter) const override {
-
-    if (!mlir::matchPattern(op.getLhs(), mlir::m_Constant()) ||
-        mlir::matchPattern(op.getRhs(), mlir::m_Constant()))
-      return mlir::failure();
-
-    using Pred = mlir::arith::CmpIPredicate;
-    const std::pair<Pred, Pred> inv[] = {
-        // clang-format off
-        {Pred::slt, Pred::sgt},
-        {Pred::sle, Pred::sge},
-        {Pred::ult, Pred::ugt},
-        {Pred::ule, Pred::uge},
-        {Pred::eq, Pred::eq},
-        {Pred::ne, Pred::ne},
-        // clang-format on
-    };
-
-    auto newPred = [&]() -> Pred {
-      auto oldPred = op.getPredicate();
-      for (auto it : inv) {
-        if (it.first == oldPred)
-          return it.second;
-        if (it.second == oldPred)
-          return it.first;
-      }
-
-      llvm_unreachable("Unknown predicate");
-    }();
-
-    rewriter.replaceOpWithNewOp<mlir::arith::CmpIOp>(op, newPred, op.getRhs(),
-                                                     op.getLhs());
-    ;
-    return mlir::success();
-  }
-};
-
 struct ReshapeAlloca : public mlir::OpRewritePattern<mlir::memref::ReshapeOp> {
   using OpRewritePattern::OpRewritePattern;
 
@@ -355,7 +313,7 @@ void PlierUtilDialect::getCanonicalizationPatterns(
   results.add<DimExpandShape<mlir::tensor::DimOp, mlir::tensor::ExpandShapeOp>,
               DimExpandShape<mlir::memref::DimOp, mlir::memref::ExpandShapeOp>,
               DimInsertSlice, FillExtractSlice, SpirvInputCSE, GenGlobalId,
-              InvertCmpi, ReshapeAlloca>(getContext());
+              ReshapeAlloca>(getContext());
 }
 
 OpaqueType OpaqueType::get(mlir::MLIRContext *context) {
@@ -1132,7 +1090,7 @@ static mlir::Value propagateCasts(mlir::Value val, mlir::Type thisType) {
 mlir::OpFoldResult SignCastOp::fold(llvm::ArrayRef<mlir::Attribute> operands) {
   assert(operands.size() == 1);
   auto thisType = getType();
-  auto attrOperand = operands.front();
+  auto attrOperand = operands.front().dyn_cast_or_null<mlir::TypedAttr>();
   if (attrOperand && attrOperand.getType() == thisType)
     return attrOperand;
 

numba_dpcomp/numba_dpcomp/mlir/tests/test_gpu.py

@@ -294,7 +294,7 @@ def func(a, b):
         b[i] = f(a[i])
 
     _test_unary(
-        func, np.float32, "GPUToSpirvPass", lambda ir: ir.count(f"OCL.{op}") == 1
+        func, np.float32, "GPUToSpirvPass", lambda ir: ir.count(f"CL.{op}") == 1
     )
 
 
@@ -647,15 +647,15 @@ def func(a, b, c, res):
         gpu_func = kernel(fastmath=False)(func)
         gpu_func[a.shape, DEFAULT_LOCAL_SIZE](a, b, c, gpu_res)
         ir = get_print_buffer()
-        assert ir.count("spv.OCL.fma") == 0, ir
+        assert ir.count("spv.CL.fma") == 0, ir
         assert_equal(gpu_res, sim_res)
 
     with print_pass_ir([], ["GPUToSpirvPass"]):
         gpu_res = np.zeros(a.shape, a.dtype)
         gpu_func = kernel(fastmath=True)(func)
         gpu_func[a.shape, DEFAULT_LOCAL_SIZE](a, b, c, gpu_res)
         ir = get_print_buffer()
-        assert ir.count("spv.OCL.fma") == 1, ir
+        assert ir.count("spv.CL.fma") == 1, ir
         assert_equal(gpu_res, sim_res)
 
 

numba_dpcomp/numba_dpcomp/mlir/tests/test_numpy.py

@@ -1217,9 +1217,6 @@ def _cov(m, y=None, rowvar=True, bias=False, ddof=None):
 _rnd = np.random.RandomState(42)
 
 
-@pytest.mark.skipif(
-    reason="Invalid folding in upstream MLIR, see https://github.com/llvm/llvm-project/issues/56557"
-)
 @parametrize_function_variants(
     "m",
     [
@@ -1259,9 +1256,6 @@ def test_cov_basic(m):
 _cov_inputs_m = _rnd.randn(105).reshape(15, 7)
 
 
-@pytest.mark.skipif(
-    reason="Invalid folding in upstream MLIR, see https://github.com/llvm/llvm-project/issues/56557"
-)
 @pytest.mark.parametrize("m", [_cov_inputs_m])
 @pytest.mark.parametrize("y", [None, _cov_inputs_m[::-1]])
 @pytest.mark.parametrize("rowvar", [False, True])
@@ -1278,9 +1272,6 @@ def test_cov_explicit_arguments(m, y, rowvar, bias, ddof):
     )
 
 
-@pytest.mark.skipif(
-    reason="Invalid folding in upstream MLIR, see https://github.com/llvm/llvm-project/issues/56557"
-)
 @parametrize_function_variants(
     "m, y, rowvar",
     [
@@ -1340,9 +1331,7 @@ def py_func(data):
     assert_equal(py_func(arr), jit_func(arr))
 
 
-@pytest.mark.skipif(
-    reason="Invalid folding in upstream MLIR, see https://github.com/llvm/llvm-project/issues/56557"
-)
+@pytest.mark.skipif(reason="Failure in propagate layout pass")
 @pytest.mark.parametrize(
     "arr",
     [

numba_dpcomp/numba_dpcomp/mlir_compiler/lib/pipelines/lower_to_llvm.cpp

@@ -816,78 +816,6 @@ struct DeallocOpLowering
   }
 };
 
-struct ReshapeLowering
-    : public mlir::ConvertOpToLLVMPattern<mlir::memref::ReshapeOp> {
-  using ConvertOpToLLVMPattern<mlir::memref::ReshapeOp>::ConvertOpToLLVMPattern;
-
-  explicit ReshapeLowering(mlir::LLVMTypeConverter &converter)
-      : ConvertOpToLLVMPattern<mlir::memref::ReshapeOp>(converter,
-                                                        /*benefit*/ 2) {}
-
-  mlir::LogicalResult
-  matchAndRewrite(mlir::memref::ReshapeOp op,
-                  mlir::memref::ReshapeOp::Adaptor adaptor,
-                  mlir::ConversionPatternRewriter &rewriter) const override {
-    auto converter = getTypeConverter();
-    auto dstType = converter->convertType(op.getType());
-    if (!dstType)
-      return mlir::failure();
-
-    mlir::MemRefDescriptor source(adaptor.source());
-    mlir::MemRefDescriptor shape(adaptor.shape());
-
-    auto loc = op.getLoc();
-    auto result = mlir::MemRefDescriptor::undef(rewriter, loc, dstType);
-    result.setAllocatedPtr(rewriter, loc, source.allocatedPtr(rewriter, loc));
-    result.setAlignedPtr(rewriter, loc, source.alignedPtr(rewriter, loc));
-    result.setOffset(rewriter, loc, source.offset(rewriter, loc));
-
-    auto memRefType = op.getType().cast<mlir::MemRefType>();
-    auto numDims = memRefType.getRank();
-    llvm::SmallVector<mlir::Value> sizes(static_cast<unsigned>(numDims));
-    auto indexType = getIndexType();
-    for (unsigned i = 0; i < numDims; ++i) {
-      auto ind = createIndexConstant(rewriter, loc, i);
-      mlir::Value dataPtr =
-          getStridedElementPtr(loc, memRefType, shape, ind, rewriter);
-      auto size = rewriter.create<mlir::LLVM::LoadOp>(loc, dataPtr).getResult();
-      if (size.getType() != indexType)
-        size = rewriter.create<mlir::LLVM::ZExtOp>(loc, indexType, size);
-
-      result.setSize(rewriter, loc, i, size);
-      sizes[i] = size;
-    }
-
-    // Strides: iterate sizes in reverse order and multiply.
-    int64_t stride = 1;
-    mlir::Value runningStride = createIndexConstant(rewriter, loc, 1);
-    for (auto i = static_cast<unsigned>(memRefType.getRank()); i-- > 0;) {
-      result.setStride(rewriter, loc, i, runningStride);
-
-      int64_t size = memRefType.getShape()[i];
-      if (size == 0)
-        continue;
-      bool useSizeAsStride = stride == 1;
-      if (size == mlir::ShapedType::kDynamicSize)
-        stride = mlir::ShapedType::kDynamicSize;
-      if (stride != mlir::ShapedType::kDynamicSize)
-        stride *= size;
-
-      if (useSizeAsStride)
-        runningStride = sizes[i];
-      else if (stride == mlir::ShapedType::kDynamicSize)
-        runningStride =
-            rewriter.create<mlir::LLVM::MulOp>(loc, runningStride, sizes[i]);
-      else
-        runningStride =
-            createIndexConstant(rewriter, loc, static_cast<uint64_t>(stride));
-    }
-
-    rewriter.replaceOp(op, static_cast<mlir::Value>(result));
-    return mlir::success();
-  }
-};
-
 struct ExpandShapeLowering
     : public mlir::ConvertOpToLLVMPattern<mlir::memref::ExpandShapeOp> {
   using ConvertOpToLLVMPattern<
@@ -1802,7 +1730,6 @@ struct PlierUtilToLLVMPass
         LowerUndef,
         LowerBuildTuple,
         LowerRetainOp,
-        ReshapeLowering,
         ExpandShapeLowering,
         LowerExtractMemrefMetadataOp,
         LowerTakeContextOp,

numba_dpcomp/numba_dpcomp/mlir_compiler/lib/pipelines/plier_to_std.cpp

@@ -302,7 +302,8 @@ static mlir::Type mapPlierType(mlir::Type type) {
 static mlir::Type dropLiteralType(mlir::Type t) {
   assert(t);
   if (auto literal = t.dyn_cast<plier::LiteralType>())
-    return dropLiteralType(literal.getValue().getType());
+    return dropLiteralType(
+        literal.getValue().cast<mlir::TypedAttr>().getType());
 
   return t;
 }
@@ -339,7 +340,8 @@ struct ConstOpLowering : public mlir::OpConversionPattern<plier::ConstOp> {
       return mlir::failure();
 
     auto value = adaptor.val();
-    if (isSupportedType(value.getType())) {
+    auto typeAttr = value.dyn_cast_or_null<mlir::TypedAttr>();
+    if (typeAttr && isSupportedType(typeAttr.getType())) {
       if (auto intAttr = value.dyn_cast<mlir::IntegerAttr>()) {
         auto type = intAttr.getType().cast<mlir::IntegerType>();
         if (!type.isSignless()) {
@@ -379,7 +381,7 @@ static bool isOmittedType(mlir::Type type) {
 }
 
 static mlir::Attribute makeSignlessAttr(mlir::Attribute val) {
-  auto type = val.getType();
+  auto type = val.cast<mlir::TypedAttr>().getType();
   if (auto intType = type.dyn_cast<mlir::IntegerType>()) {
     if (!intType.isSignless()) {
       auto newType = plier::makeSignlessType(intType);
@@ -411,8 +413,8 @@ struct LiteralLowering : public mlir::OpConversionPattern<Op> {
     if (auto literal = convertedType.template dyn_cast<plier::LiteralType>()) {
       auto loc = op.getLoc();
       auto attrVal = literal.getValue();
-      auto dstType = attrVal.getType();
-      auto val = makeSignlessAttr(attrVal);
+      auto dstType = attrVal.template cast<mlir::TypedAttr>().getType();
+      auto val = makeSignlessAttr(attrVal).template cast<mlir::TypedAttr>();
       auto newVal =
           rewriter.create<mlir::arith::ConstantOp>(loc, val).getResult();
       if (dstType != val.getType())
@@ -470,11 +472,11 @@ struct OmittedLowering : public mlir::OpConversionPattern<plier::CastOp> {
     if (auto omittedAttr =
             getOmittedValue(adaptor.value().getType(), convertedType)) {
       auto loc = op.getLoc();
-      auto dstType = omittedAttr.getType();
+      auto dstType = omittedAttr.cast<mlir::TypedAttr>().getType();
       auto val = makeSignlessAttr(omittedAttr);
       auto newVal =
           rewriter.create<mlir::arith::ConstantOp>(loc, val).getResult();
-      if (dstType != val.getType())
+      if (dstType != val.cast<mlir::TypedAttr>().getType())
         newVal = rewriter.create<plier::SignCastOp>(loc, dstType, newVal);
 
       rewriter.replaceOp(op, newVal);
@@ -697,8 +699,8 @@ mlir::Value doCast(mlir::PatternRewriter &rewriter, mlir::Location loc,
   assert(dstType);
   auto srcType = val.getType();
   if (auto literal = srcType.dyn_cast<plier::LiteralType>()) {
-    auto attr = literal.getValue();
-    auto signlessAttr = makeSignlessAttr(attr);
+    auto attr = literal.getValue().cast<mlir::TypedAttr>();
+    auto signlessAttr = makeSignlessAttr(attr).cast<mlir::TypedAttr>();
     val = rewriter.create<mlir::arith::ConstantOp>(loc, signlessAttr);
     if (signlessAttr.getType() != attr.getType())
       val = rewriter.create<plier::SignCastOp>(loc, attr.getType(), val);
@@ -1413,7 +1415,7 @@ void PlierToStdPass::runOnOperation() {
           return false;
 
         if (auto literal = type.dyn_cast<plier::LiteralType>())
-          type = literal.getValue().getType();
+          type = literal.getValue().cast<mlir::TypedAttr>().getType();
 
         return !type.isIntOrFloat();
       });
@@ -1462,8 +1464,9 @@ struct ConvertLiteralTypesPass
     typeConverter.addConversion([](mlir::Type type) { return type; });
 
     auto context = &getContext();
-    typeConverter.addConversion(
-        [](plier::LiteralType type) { return type.getValue().getType(); });
+    typeConverter.addConversion([](plier::LiteralType type) {
+      return type.getValue().cast<mlir::TypedAttr>().getType();
+    });
 
     auto materializeCast =
         [](mlir::OpBuilder &builder, mlir::Type type, mlir::ValueRange inputs,

numba_dpcomp/numba_dpcomp/mlir_compiler/lib/py_linalg_resolver.cpp

@@ -16,6 +16,7 @@
 
 #include <pybind11/pybind11.h>
 
+#include <mlir/AsmParser/AsmParser.h>
 #include <mlir/Dialect/Arithmetic/IR/Arithmetic.h>
 #include <mlir/Dialect/Bufferization/IR/Bufferization.h>
 #include <mlir/Dialect/Func/IR/FuncOps.h>
@@ -220,7 +221,9 @@ static auto toValues(py::handle obj, UnwrapFunc &&unwrapFunc) {
 static llvm::Optional<py::object> getPyLiteral(mlir::Attribute attr) {
   assert(attr);
   if (auto intAttr = attr.dyn_cast<mlir::IntegerAttr>()) {
-    if (auto intType = attr.getType().dyn_cast<mlir::IntegerType>()) {
+    if (auto intType = attr.cast<mlir::TypedAttr>()
+                           .getType()
+                           .dyn_cast<mlir::IntegerType>()) {
       // Ignore index type
       if (intType.getWidth() == 1)
         return py::bool_(intAttr.getInt() != 0);

numba_dpcomp/setup.py

@@ -66,6 +66,7 @@
         "-DIMEX_ENABLE_NUMBA_FE=ON",
         "-DIMEX_ENABLE_NUMBA_HOTFIX=ON",
         "-DIMEX_ENABLE_TBB_SUPPORT=ON",
+        "-DLLVM_ENABLE_ZSTD=OFF",
     ]
 
     # DPNP