Upstream merge 55

external/mlir-hal/lib/Dialect/MHAL/Transforms/BufferizableOpInterfaceImpl.cpp

@@ -10,11 +10,13 @@
 
 #include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h"
 #include "mlir/Dialect/Bufferization/IR/Bufferization.h"
+#include "mlir/Dialect/Bufferization/IR/BufferizationTypeInterfaces.h"
 #include "mlir/Dialect/Bufferization/Transforms/FuncBufferizableOpInterfaceImpl.h"
 #include "mlir/Dialect/Bufferization/Transforms/OneShotAnalysis.h"
 #include "mlir/Dialect/Func/IR/FuncOps.h"
 #include "mlir/Dialect/MHAL/IR/MHAL.h"
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/Dialect.h"
 #include "mlir/IR/Operation.h"
 #include "mlir/IR/PatternMatch.h"
@@ -124,11 +126,13 @@ struct LaunchOpInterface
       Type returnType = returnVal.getType();
       if (isa<TensorType>(returnType)) {
         assert(returnType == callResultTypes[funcResultIdx++]);
-        FailureOr<BaseMemRefType> memrefType =
+        FailureOr<BufferLikeType> bufferType =
             bufferization::getBufferType(returnVal, options, state);
-        if (failed(memrefType))
+        if (failed(bufferType))
           return failure();
-        resultTypes.push_back(*memrefType);
+        assert(isa<BaseMemRefType>(*bufferType) && "expected memref type");
+        BaseMemRefType memrefType = cast<BaseMemRefType>(*bufferType);
+        resultTypes.push_back(memrefType);
       } else {
         // Non-tensor values are returned.
         resultTypes.push_back(returnType);

mlir/lib/Conversion/EmulateFp8ExtTrunc/EmulateFp8ExtTrunc.cpp

@@ -311,14 +311,14 @@ static FlatSymbolRefAttr makeFp8TruncFunction(Location loc, FloatType outType,
   Value cmp = b.create<CmpIOp>(CmpIPredicate::eq, and4, infNanConst);
 
   Block *notInfNan = func.addBlock();
-  Value outNan = b.create<ConstantFloatOp>(APFloat::getQNaN(outSem), outType);
+  Value outNan = b.create<ConstantFloatOp>(outType, APFloat::getQNaN(outSem));
   b.create<cf::CondBranchOp>(cmp, ret, ValueRange{outNan}, notInfNan,
                              ValueRange{});
   b.setInsertionPointToStart(notInfNan);
 
   // A deviation from the MIGraphX: denormals are zero here
   Value cmp5 = b.create<CmpIOp>(CmpIPredicate::eq, and2, i32Const(0));
-  Value outZero = b.create<ConstantFloatOp>(APFloat::getZero(outSem), outType);
+  Value outZero = b.create<ConstantFloatOp>(outType, APFloat::getZero(outSem));
   Block *notZero = func.addBlock();
   b.create<cf::CondBranchOp>(cmp5, ret, ValueRange{outZero}, notZero,
                              ValueRange{});
@@ -366,7 +366,7 @@ static FlatSymbolRefAttr makeFp8TruncFunction(Location loc, FloatType outType,
   Value cmp57 = b.create<CmpIOp>(CmpIPredicate::ne, sub43, i32Const(0));
   Value and58 = b.create<AndIOp>(add56, i32Const(1 << 23));
   Value tobool59Not = b.create<CmpIOp>(CmpIPredicate::eq, and58, i32Const(0));
-  Value trueConst = b.create<ConstantIntOp>(true, 1);
+  Value trueConst = b.create<ConstantIntOp>(true, /*width=*/1);
   Value brCond133 = b.create<SelectOp>(cmp57, trueConst, tobool59Not);
 
   Block *ifElse61 = func.addBlock();
@@ -392,7 +392,7 @@ static FlatSymbolRefAttr makeFp8TruncFunction(Location loc, FloatType outType,
   b.setInsertionPointToStart(ifThen70);
   Value ir5 = b.create<TruncIOp>(i8, ir1);
   Value conv =
-      b.create<OrIOp>(ir5, b.create<ConstantIntOp>(127, b.getI8Type()));
+      b.create<OrIOp>(ir5, b.create<ConstantIntOp>(b.getI8Type(), 127));
   Value convOut = b.create<BitcastOp>(outType, conv);
   b.create<cf::BranchOp>(ret, convOut);
 
@@ -402,7 +402,7 @@ static FlatSymbolRefAttr makeFp8TruncFunction(Location loc, FloatType outType,
   Value cmp72 = b.create<CmpIOp>(CmpIPredicate::eq, f8Exponent0, i32Const(0));
   Value cmp74 = b.create<CmpIOp>(CmpIPredicate::ult, mantissa1,
                                  i32Const(1 << (16 + eBits)));
-  Value falseConst = b.create<ConstantIntOp>(false, 1);
+  Value falseConst = b.create<ConstantIntOp>(false, /*width=*/1);
   Value brCond = b.create<SelectOp>(cmp72, cmp74, falseConst);
   b.create<cf::CondBranchOp>(brCond, ret, ValueRange{outZero}, ifEnd76,
                              ValueRange{f8Exponent0, mantissa1});
@@ -735,7 +735,7 @@ void Fp8TruncToCallPattern::rewrite(TruncFOp op, OpAdaptor adaptor,
   Value rets = rewriter.createOrFold<vector::SplatOp>(
       loc,
       rewriter.createOrFold<ConstantFloatOp>(
-          loc, APFloat::getZero(outElemType.getFloatSemantics()), outElemType),
+          loc, outElemType, APFloat::getZero(outElemType.getFloatSemantics())),
       retVecType);
   SmallVector<int64_t> strides = computeStrides(inVecType.getShape());
   for (int64_t i = 0, e = inVecType.getNumElements(); i < e; ++i) {

mlir/lib/Dialect/Rock/Transforms/SugarToLoops.cpp

@@ -916,7 +916,7 @@ static Value getConstIntOrIndexValue(OpBuilder &b, Location loc, int64_t value,
   if (isa<IndexType>(type)) {
     return b.create<ConstantIndexOp>(loc, value);
   }
-  return b.create<ConstantIntOp>(loc, value, type);
+  return b.create<ConstantIntOp>(loc, type, value);
 }
 
 // Manually flatten a set of coordinates into a single address

mlir/lib/Dialect/Rock/utility/builderUtils.cpp

@@ -209,7 +209,7 @@ Value getAsTensor(OpBuilder &builder, Location loc, mlir::Value value,
                   bool isWritable) {
   constexpr bool isRestrict{true};
   Value origTensor = builder.create<bufferization::ToTensorOp>(
-      loc, value, isRestrict, isWritable);
+      loc, value.getType(), value, isRestrict, isWritable);
   return origTensor;
 }
 

mlir/tools/rocmlir-gen/rocmlir-gen.cpp

@@ -3280,7 +3280,8 @@ createCpuConvElementwiseGemmKernelWithMlir(ModuleOp module,
                                           bool isWritable = false) {
     constexpr bool isRestrict{true};
     Value flatTensor = builder.create<bufferization::ToTensorOp>(
-        loc, block->getArgument(blockArgIndex), isRestrict, isWritable);
+        loc, block->getArgument(blockArgIndex).getType(),
+        block->getArgument(blockArgIndex), isRestrict, isWritable);
     ArrayRef<int64_t> origShape =
         cast<ShapedType>(argTypes[blockArgIndex]).getShape();
 
@@ -3459,7 +3460,8 @@ createCpuGemmElementwiseGemmKernelWithMlir(ModuleOp module,
                                           bool isWritable = false) {
     constexpr bool isRestrict{true};
     Value flatTensor = builder.create<bufferization::ToTensorOp>(
-        loc, block->getArgument(blockArgIndex), isRestrict, isWritable);
+        loc, block->getArgument(blockArgIndex).getType(),
+        block->getArgument(blockArgIndex), isRestrict, isWritable);
     ArrayRef<int64_t> origShape =
         cast<ShapedType>(argTypes[blockArgIndex]).getShape();
 
@@ -3574,7 +3576,8 @@ static func::FuncOp createCpuAttentionKernelWithMlir(ModuleOp module,
                                           bool isWritable = false) {
     constexpr bool isRestrict{true};
     Value flatTensor = builder.create<bufferization::ToTensorOp>(
-        loc, block->getArgument(blockArgIndex), isRestrict, isWritable);
+        loc, block->getArgument(blockArgIndex).getType(),
+        block->getArgument(blockArgIndex), isRestrict, isWritable);
     ArrayRef<int64_t> origShape =
         cast<ShapedType>(argTypes[blockArgIndex]).getShape();
 
@@ -3935,7 +3938,7 @@ static func::FuncOp createVerifierFunc(ModuleOp module, const KernelIF &kernel,
   char printDebug = static_cast<char>(printVerifyResults.getValue());
 
   auto printDebugVal =
-      b.create<arith::ConstantIntOp>(loc, printDebug, charType);
+      b.create<arith::ConstantIntOp>(loc, charType, printDebug);
 
   // obtain function name of the verifier wrapper
   std::string verifyFuncName = "mcpuVerify";