[Mlir-commits] [mlir] [MLIR][Linalg] Diagnose unsupported types in Linalg named op region builders (PR #181616)

[llvmlistbot at llvm.org]

================
@@ -6582,13 +6589,18 @@ void BatchReduceMatmulOp::regionBuilder(
   SmallVector<Value> yields;
 
   auto toType = block.getArgument(2).getType();
-  Value castValA =
-      helper.buildTypeFn(TypeFn::cast_signed, toType, block.getArgument(0));
-  Value castValB =
-      helper.buildTypeFn(TypeFn::cast_signed, toType, block.getArgument(1));
-  Value mulVal = helper.buildBinaryFn(BinaryFn::mul, castValA, castValB);
+  Value castValA = helper.buildTypeFn(TypeFn::cast_signed, toType,
+                                      block.getArgument(0), emitError);
+  Value castValB = helper.buildTypeFn(TypeFn::cast_signed, toType,
+                                      block.getArgument(1), emitError);
+  Value mulVal =
+      helper.buildBinaryFn(BinaryFn::mul, castValA, castValB, emitError);
+  if (!castValA || !castValB || !mulVal)
+    return;
   Value addVal =
       helper.buildBinaryFn(BinaryFn::add, block.getArgument(2), mulVal);
+  if (!addVal)
----------------
shubhamnarlawar wrote:

In case of amx.tile as operand which is not supported, mulVal is null and hence if I pass null value to add, it reaches llvm_unreachable() in buildBinaryFn() and crashes as below -

`invalid.mlir:4:35: error: custom op 'linalg.batch_reduce_matmul' Cannot build binary Linalg operation: expects allComplex, allFloatingPoint, or allInteger, got '!amx.tile<16x16xbf16>' and '!amx.tile<16x16xbf16>'
  %1 = linalg.batch_reduce_matmul ins(%0, %0 : !amx.tile<16x16xbf16>, !amx.tile<16x16xbf16>) outs(%0 : !amx.tile<16x16xbf16>) -> !amx.tile<16x16xbf16>
                                  ^
unsupported non numeric type
UNREACHABLE executed at /home/shubham/work/llvm-project/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp:516!

PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace and instructions to reproduce the bug.
Stack dump:
0.  Program arguments: ../build/bin/mlir-opt invalid.mlir
1.  MLIR Parser: custom op parser 'func.func'
2.  MLIR Parser: custom op parser 'linalg.batch_reduce_matmul'`

https://github.com/llvm/llvm-project/pull/181616