[MLIR] Error handling in MaterializeVectors

This removes assertions as a means to capture NYI behavior and propagates
errors up.

PiperOrigin-RevId: 224376935

1 files changed, 41 insertions, 16 deletions

diff --git a/mlir/lib/Transforms/MaterializeVectors.cpp b/mlir/lib/Transforms/MaterializeVectors.cpp
index 7cd6a0ad273..3bf4305ca0c 100644
--- a/mlir/lib/Transforms/MaterializeVectors.cpp
+++ b/mlir/lib/Transforms/MaterializeVectors.cpp
@@ -209,23 +209,25 @@ struct MaterializeVectors : public FunctionPass {
 
 char MaterializeVectors::passID = 0;
 
-// Returns the distance, in number of elements, between a slice in a dimension
-// and the next slice in the same dimension.
-//   e.g. shape[3, 4, 5] -> strides[20, 5, 1]
+/// Given a shape with sizes greater than 0 along all dimensions,
+/// returns the distance, in number of elements, between a slice in a dimension
+/// and the next slice in the same dimension.
+///   e.g. shape[3, 4, 5] -> strides[20, 5, 1]
 static SmallVector<unsigned, 8> makeStrides(ArrayRef<unsigned> shape) {
   SmallVector<unsigned, 8> tmp;
   tmp.reserve(shape.size());
   unsigned running = 1;
   for (auto rit = shape.rbegin(), reit = shape.rend(); rit != reit; ++rit) {
-    // TODO(ntv): emitError instead of NYI assert.
-    assert(*rit > 0 && "NYI: symbolic or null shape dimension");
+    assert(*rit > 0 && "size must be greater than 0 along all dimensions of "
+                       "shape");
     tmp.push_back(running);
     running *= *rit;
   }
   return SmallVector<unsigned, 8>(tmp.rbegin(), tmp.rend());
 }
 
-// Returns the linearized expression.
+/// Given a shape with sizes greater than 0 along all dimensions, returns the
+/// delinearized components of linearIndex along shape.
 static SmallVector<unsigned, 8> delinearize(unsigned linearIndex,
                                             ArrayRef<unsigned> shape) {
   SmallVector<unsigned, 8> res;
@@ -256,6 +258,8 @@ instantiate(MLFuncBuilder *b, OperationStmt *opStmt, VectorType hwVectorType,
 /// insertion.
 /// For now, this is limited to ConstantOp because we do not vectorize loop
 /// indices and will need to be extended in the future.
+///
+/// If substitution fails, returns nullptr.
 static MLValue *
 substitute(SSAValue *v, VectorType hwVectorType,
            DenseMap<const MLValue *, MLValue *> *substitutionsMap) {
@@ -271,7 +275,7 @@ substitute(SSAValue *v, VectorType hwVectorType,
       return res.first->second;
     }
     v->getDefiningOperation()->emitError("Missing substitution");
-    assert(false);
+    return nullptr;
   }
   return it->second;
 }
@@ -400,6 +404,8 @@ materializeAttributes(OperationStmt *opStmt, VectorType hwVectorType) {
 /// affine reindexing. Just substitute their SSAValue* operands and be done. For
 /// this case the actual instance is irrelevant. Just use the SSA values in
 /// substitutionsMap.
+///
+/// If the underlying substitution fails, this fails too and returns nullptr.
 static OperationStmt *
 instantiate(MLFuncBuilder *b, OperationStmt *opStmt, VectorType hwVectorType,
             DenseMap<const MLValue *, MLValue *> *substitutionsMap) {
@@ -407,11 +413,18 @@ instantiate(MLFuncBuilder *b, OperationStmt *opStmt, VectorType hwVectorType,
          "Should call the function specialized for VectorTransferReadOp");
   assert(!opStmt->isa<VectorTransferWriteOp>() &&
          "Should call the function specialized for VectorTransferWriteOp");
+  bool fail = false;
   auto operands = map(
-      [hwVectorType, substitutionsMap](SSAValue *v) {
-        return substitute(v, hwVectorType, substitutionsMap);
+      [hwVectorType, substitutionsMap, &fail](SSAValue *v) {
+        auto *res =
+            fail ? nullptr : substitute(v, hwVectorType, substitutionsMap);
+        fail |= !res;
+        return res;
       },
       opStmt->getOperands());
+  if (fail) {
+    return nullptr;
+  }
   auto attrs = materializeAttributes(opStmt, hwVectorType);
   return b->createOperation(opStmt->getLoc(), opStmt->getName(), operands,
                             {hwVectorType}, attrs);
@@ -452,7 +465,7 @@ static AffineMap projectedPermutationMap(VectorTransferOpTy *transfer,
   auto permutationMap = transfer->getPermutationMap();
   LLVM_DEBUG(projectionMap.print(dbgs() << "\nprojectionMap: "));
   LLVM_DEBUG(permutationMap.print(dbgs() << "\npermutationMap: "));
-  return composeUnboundedMaps(projectionMap, transfer->getPermutationMap());
+  return composeUnboundedMaps(projectionMap, permutationMap);
 }
 
 /// Creates an instantiated version of `read` for the instance of
@@ -516,6 +529,8 @@ instantiate(MLFuncBuilder *b, VectorTransferWriteOp *write,
 ///      type, all operands are substituted according to `substitutions`. Thanks
 ///      to the topological order of a slice, the substitution is always
 ///      possible.
+///
+/// Returns true on failure.
 static bool instantiateMaterialization(Statement *stmt,
                                        MaterializationState *state) {
   LLVM_DEBUG(dbgs() << "\ninstantiate: " << *stmt);
@@ -557,6 +572,9 @@ static bool instantiateMaterialization(Statement *stmt,
   }
   auto *clone =
       instantiate(&b, opStmt, state->hwVectorType, state->substitutionsMap);
+  if (!clone) {
+    return true;
+  }
   state->substitutionsMap->insert(std::make_pair(
       cast<MLValue>(opStmt->getResult(0)), cast<MLValue>(clone->getResult(0))));
   return false;
@@ -578,10 +596,12 @@ static bool instantiateMaterialization(Statement *stmt,
 ///      equivalent of loop strip-mining + loop sinking and encoded this in the
 ///      vector type.
 ///
+/// Returns true on failure.
+///
 /// TODO(ntv): materialized allocs.
 /// TODO(ntv): full loops + materialized allocs.
 /// TODO(ntv): partial unrolling + materialized allocs.
-static void emitSlice(MaterializationState *state,
+static bool emitSlice(MaterializationState *state,
                       SetVector<Statement *> *slice) {
   auto ratio = shapeRatio(state->superVectorType, state->hwVectorType);
   assert(ratio.hasValue() &&
@@ -601,7 +621,7 @@ static void emitSlice(MaterializationState *state,
       auto fail = instantiateMaterialization(stmt, &scopedState);
       if (fail) {
         stmt->emitError("Unhandled super-vector materialization failure");
-        assert(!fail);
+        return true;
       }
     }
   }
@@ -618,6 +638,7 @@ static void emitSlice(MaterializationState *state,
     LLVM_DEBUG((*slice)[idx]->print(dbgs()));
     (*slice)[idx]->erase();
   }
+  return false;
 }
 
 /// Materializes super-vector types into concrete hw vector types as follows:
@@ -637,7 +658,7 @@ static void emitSlice(MaterializationState *state,
 /// Additionally, this set is limited to statements in the same lexical scope
 /// because we currently disallow vectorization of defs that come from another
 /// scope.
-static void materialize(MLFunction *f,
+static bool materialize(MLFunction *f,
                         const SetVector<OperationStmt *> &terminators,
                         MaterializationState *state) {
   DenseSet<Statement *> seen;
@@ -686,10 +707,14 @@ static void materialize(MLFunction *f,
            "Only f32 supported for now");
     state->hwVectorType = VectorType::get(
         state->hwVectorSize, state->superVectorType.getElementType());
-    emitSlice(state, &slice);
+    auto fail = emitSlice(state, &slice);
+    if (fail) {
+      return true;
+    }
     LLVM_DEBUG(dbgs() << "\nMLFunction is now\n");
     LLVM_DEBUG(f->print(dbgs()));
   }
+  return false;
 }
 
 PassResult MaterializeVectors::runOnMLFunction(MLFunction *f) {
@@ -720,9 +745,9 @@ PassResult MaterializeVectors::runOnMLFunction(MLFunction *f) {
   }
 
   // Call materialization.
-  materialize(f, terminators, &state);
+  auto fail = materialize(f, terminators, &state);
 
-  return PassResult::Success;
+  return fail ? PassResult::Failure : PassResult::Success;
 }
 
 FunctionPass *mlir::createMaterializeVectors() {