[MLIR][Affine] Fix affine.apply verifier and add functionality to demote invalid symbols to dims (#128289)

Fixes: #120189,
#128403

Fix affine.apply verifier to reject symbolic operands that are valid
dims for affine purposes. This doesn't affect other users in other
contexts where the operands were neither valid dims or symbols (for eg.
in scf.for or other region ops). Otherwise, it was possible for
`-canonicalize` to have generated invalid IR when such
affine.apply ops were composed.

Introduce a method to demote a symbolic operand to a dimensional one
(the inverse of the current canonicalizePromotedSymbols).  Demote
operands that could/should have been valid affine dimensional values
(affine loop IVs or their functions) from symbols to dims. This is a
general method that can be used to legalize a map + operands post
construction depending on its operands. Use it during
`canonicalizeMapOrSetAndOperands` so that pattern rewriter-based passes
are able to generate valid IR post folding. Users outside of affine
analyses/dialects remain unaffected.

In some cases, this change also leads to better simplified operands,
duplicates eliminated as shown in one of the test cases where the same
operand appeared as a symbol and as a dim.

This commit also fixes test cases where dimensional positions should have
been ideally used with affine.apply (for affine loop IVs for example).

Author: bondhugula

mlir/include/mlir/Dialect/Affine/IR/AffineOps.td

@@ -40,21 +40,25 @@ def AffineApplyOp : Affine_Op<"apply", [Pure]> {
   let description = [{
     The `affine.apply` operation applies an [affine mapping](#affine-maps)
     to a list of SSA values, yielding a single SSA value. The number of
-    dimension and symbol arguments to `affine.apply` must be equal to the
+    dimension and symbol operands to `affine.apply` must be equal to the
     respective number of dimensional and symbolic inputs to the affine mapping;
     the affine mapping has to be one-dimensional, and so the `affine.apply`
     operation always returns one value. The input operands and result must all
     have ‘index’ type.
 
+    An operand that is a valid dimension as per the [rules on valid affine
+    dimensions and symbols](#restrictions-on-dimensions-and-symbols)
+    cannot be used as a symbolic operand.
+
     Example:
 
     ```mlir
-    #map10 = affine_map<(d0, d1) -> (d0 floordiv 8 + d1 floordiv 128)>
+    #map = affine_map<(d0, d1) -> (d0 floordiv 8 + d1 floordiv 128)>
     ...
-    %1 = affine.apply #map10 (%s, %t)
+    %1 = affine.apply #map (%s, %t)
 
     // Inline example.
-    %2 = affine.apply affine_map<(i)[s0] -> (i+s0)> (%42)[%n]
+    %2 = affine.apply affine_map<(i)[s0] -> (i + s0)> (%42)[%n]
     ```
   }];
   let arguments = (ins AffineMapAttr:$map, Variadic<Index>:$mapOperands);

mlir/lib/Dialect/Affine/IR/AffineOps.cpp

@@ -578,6 +578,15 @@ LogicalResult AffineApplyOp::verify() {
   if (affineMap.getNumResults() != 1)
     return emitOpError("mapping must produce one value");
 
+  // Do not allow valid dims to be used in symbol positions. We do allow
+  // affine.apply to use operands for values that may neither qualify as affine
+  // dims or affine symbols due to usage outside of affine ops, analyses, etc.
+  Region *region = getAffineScope(*this);
+  for (Value operand : getMapOperands().drop_front(affineMap.getNumDims())) {
+    if (::isValidDim(operand, region) && !::isValidSymbol(operand, region))
+      return emitError("dimensional operand cannot be used as a symbol");
+  }
+
   return success();
 }
 
@@ -1359,13 +1368,64 @@ static void canonicalizePromotedSymbols(MapOrSet *mapOrSet,
 
   resultOperands.append(remappedSymbols.begin(), remappedSymbols.end());
   *operands = resultOperands;
-  *mapOrSet = mapOrSet->replaceDimsAndSymbols(dimRemapping, {}, nextDim,
-                                              oldNumSyms + nextSym);
+  *mapOrSet = mapOrSet->replaceDimsAndSymbols(
+      dimRemapping, /*symReplacements=*/{}, nextDim, oldNumSyms + nextSym);
 
   assert(mapOrSet->getNumInputs() == operands->size() &&
          "map/set inputs must match number of operands");
 }
 
+/// A valid affine dimension may appear as a symbol in affine.apply operations.
+/// Given an application of `operands` to an affine map or integer set
+/// `mapOrSet`, this function canonicalizes symbols of `mapOrSet` that are valid
+/// dims, but not valid symbols into actual dims. Without such a legalization,
+/// the affine.apply will be invalid. This method is the exact inverse of
+/// canonicalizePromotedSymbols.
+template <class MapOrSet>
+static void legalizeDemotedDims(MapOrSet &mapOrSet,
+                                SmallVectorImpl<Value> &operands) {
+  if (!mapOrSet || operands.empty())
+    return;
+
+  unsigned numOperands = operands.size();
+
+  assert(mapOrSet->getNumInputs() == numOperands &&
+         "map/set inputs must match number of operands");
+
+  auto *context = mapOrSet.getContext();
+  SmallVector<Value, 8> resultOperands;
+  resultOperands.reserve(numOperands);
+  SmallVector<Value, 8> remappedDims;
+  remappedDims.reserve(numOperands);
+  SmallVector<Value, 8> symOperands;
+  symOperands.reserve(mapOrSet.getNumSymbols());
+  unsigned nextSym = 0;
+  unsigned nextDim = 0;
+  unsigned oldNumDims = mapOrSet.getNumDims();
+  SmallVector<AffineExpr, 8> symRemapping(mapOrSet.getNumSymbols());
+  resultOperands.assign(operands.begin(), operands.begin() + oldNumDims);
+  for (unsigned i = oldNumDims, e = mapOrSet.getNumInputs(); i != e; ++i) {
+    if (operands[i] && isValidDim(operands[i]) && !isValidSymbol(operands[i])) {
+      // This is a valid dim that appears as a symbol, legalize it.
+      symRemapping[i - oldNumDims] =
+          getAffineDimExpr(oldNumDims + nextDim++, context);
+      remappedDims.push_back(operands[i]);
+    } else {
+      symRemapping[i - oldNumDims] = getAffineSymbolExpr(nextSym++, context);
+      symOperands.push_back(operands[i]);
+    }
+  }
+
+  append_range(resultOperands, remappedDims);
+  append_range(resultOperands, symOperands);
+  operands = resultOperands;
+  mapOrSet = mapOrSet.replaceDimsAndSymbols(
+      /*dimReplacements=*/{}, symRemapping, oldNumDims + nextDim, nextSym);
+
+  assert(mapOrSet->getNumInputs() == operands.size() &&
+         "map/set inputs must match number of operands");
+}
+
 // Works for either an affine map or an integer set.
 template <class MapOrSet>
 static void canonicalizeMapOrSetAndOperands(MapOrSet *mapOrSet,
@@ -1380,6 +1440,7 @@ static void canonicalizeMapOrSetAndOperands(MapOrSet *mapOrSet,
          "map/set inputs must match number of operands");
 
   canonicalizePromotedSymbols<MapOrSet>(mapOrSet, operands);
+  legalizeDemotedDims<MapOrSet>(*mapOrSet, *operands);
 
   // Check to see what dims are used.
   llvm::SmallBitVector usedDims(mapOrSet->getNumDims());

mlir/test/Dialect/Affine/canonicalize.mlir

@@ -1460,8 +1460,8 @@ func.func @mod_of_mod(%lb: index, %ub: index, %step: index) -> (index, index) {
 func.func @prefetch_canonicalize(%arg0: memref<512xf32>) -> () {
   // CHECK: affine.for [[I_0_:%.+]] = 0 to 8 {
   affine.for %arg3 = 0 to 8  {
-    %1 = affine.apply affine_map<()[s0] -> (s0 * 64)>()[%arg3]
-    // CHECK: affine.prefetch [[PARAM_0_]][symbol([[I_0_]]) * 64], read, locality<3>, data : memref<512xf32>
+    %1 = affine.apply affine_map<(d0) -> (d0 * 64)>(%arg3)
+    // CHECK: affine.prefetch [[PARAM_0_]][[[I_0_]] * 64], read, locality<3>, data : memref<512xf32>
     affine.prefetch %arg0[%1], read, locality<3>, data : memref<512xf32>
   }
   return

mlir/test/Dialect/Affine/invalid.mlir

@@ -563,3 +563,17 @@ func.func @no_upper_bound() {
   }
   return
 }
+
+// -----
+
+func.func @invalid_symbol() {
+  affine.for %arg1 = 0 to 1 {
+    affine.for %arg2 = 0 to 26 {
+      affine.for %arg3 = 0 to 23 {
+        affine.apply affine_map<()[s0, s1] -> (s0 * 23 + s1)>()[%arg1, %arg3]
+        // expected-error@above {{dimensional operand cannot be used as a symbol}}
+      }
+    }
+  }
+  return
+}

mlir/test/Dialect/MemRef/fold-memref-alias-ops.mlir

@@ -496,8 +496,8 @@ func.func @fold_dynamic_subview_with_memref_store_expand_shape(%arg0 : memref<16
 
 // -----
 
-// CHECK-DAG: #[[$MAP0:.*]] = affine_map<()[s0, s1] -> (s0 + s1)>
-// CHECK-DAG: #[[$MAP1:.*]] = affine_map<()[s0] -> (s0 * 3)>
+// CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0)[s0] -> (d0 + s0)>
+// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0) -> (d0 * 3)>
 // CHECK-LABEL: fold_memref_alias_expand_shape_subview_load_store_dynamic_dim
 // CHECK-SAME: (%[[ARG0:.*]]: memref<2048x16xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: index, %[[ARG3:.*]]: index, %[[ARG4:.*]]: index)
 func.func @fold_memref_alias_expand_shape_subview_load_store_dynamic_dim(%alloc: memref<2048x16xf32>, %c10: index, %c5: index, %c0: index, %sz0: index) {
@@ -518,16 +518,16 @@ func.func @fold_memref_alias_expand_shape_subview_load_store_dynamic_dim(%alloc:
 // CHECK-NEXT:   %[[DIM:.*]] = memref.dim %[[EXPAND_SHAPE]], %[[ARG3]] : memref<?x1x8x2xf32, strided<[16, 16, 2, 1], offset: ?>>
 // CHECK-NEXT:   affine.for %[[ARG4:.*]] = 0 to %[[DIM]] step 64 {
 // CHECK-NEXT:   affine.for %[[ARG5:.*]] = 0 to 16 step 16 {
-// CHECK-NEXT:   %[[VAL0:.*]] = affine.apply #[[$MAP0]]()[%[[ARG2]], %[[ARG4]]]
-// CHECK-NEXT:   %[[VAL1:.*]] = affine.apply #[[$MAP1]]()[%[[ARG5]]]
+// CHECK-NEXT:   %[[VAL0:.*]] = affine.apply #[[$MAP0]](%[[ARG4]])[%[[ARG2]]]
+// CHECK-NEXT:   %[[VAL1:.*]] = affine.apply #[[$MAP1]](%[[ARG5]])
 // CHECK-NEXT:   %[[VAL2:.*]] = affine.load %[[ARG0]][%[[VAL0]], %[[VAL1]]] : memref<2048x16xf32>
-// CHECK-NEXT:   %[[VAL3:.*]] = affine.apply #[[$MAP0]]()[%[[ARG2]], %[[ARG4]]]
+// CHECK-NEXT:   %[[VAL3:.*]] = affine.apply #[[$MAP0]](%[[ARG4]])[%[[ARG2]]]
 // CHECK-NEXT:   affine.store %[[VAL2]], %[[ARG0]][%[[VAL3]], %[[ARG5]]] : memref<2048x16xf32>
 
 // -----
 
-// CHECK-DAG: #[[$MAP0:.*]] = affine_map<()[s0, s1] -> (s0 * 1024 + s1)>
-// CHECK-DAG: #[[$MAP1:.*]] = affine_map<()[s0, s1] -> (s0 + s1)>
+// CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0 * 1024 + d1)>
+// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d0 + d1)>
 // CHECK-LABEL: fold_static_stride_subview_with_affine_load_store_expand_shape
 // CHECK-SAME: (%[[ARG0:.*]]: memref<1024x1024xf32>, %[[ARG1:.*]]: memref<1xf32>, %[[ARG2:.*]]: index)
 func.func @fold_static_stride_subview_with_affine_load_store_expand_shape(%arg0: memref<1024x1024xf32>, %arg1: memref<1xf32>, %arg2: index) -> f32 {
@@ -549,14 +549,14 @@ func.func @fold_static_stride_subview_with_affine_load_store_expand_shape(%arg0:
 // CHECK-NEXT:  affine.for %[[ARG4:.*]] = 0 to 1024 {
 // CHECK-NEXT:   affine.for %[[ARG5:.*]] = 0 to 1020 {
 // CHECK-NEXT:    affine.for %[[ARG6:.*]] = 0 to 1 {
-// CHECK-NEXT:     %[[IDX1:.*]] = affine.apply #[[$MAP0]]()[%[[ARG3]], %[[ARG4]]]
-// CHECK-NEXT:     %[[IDX2:.*]] = affine.apply #[[$MAP1]]()[%[[ARG5]], %[[ARG6]]]
+// CHECK-NEXT:     %[[IDX1:.*]] = affine.apply #[[$MAP0]](%[[ARG3]], %[[ARG4]])
+// CHECK-NEXT:     %[[IDX2:.*]] = affine.apply #[[$MAP1]](%[[ARG5]], %[[ARG6]])
 // CHECK-NEXT:     affine.load %[[ARG0]][%[[IDX1]], %[[IDX2]]] : memref<1024x1024xf32>
 
 // -----
 
-// CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1)[s0] -> (d0 + d1 + s0 * 1024)>
-// CHECK-DAG: #[[$MAP1:.*]] = affine_map<()[s0, s1] -> (s0 + s1)>
+// CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0 * 1025 + d1)>
+// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d0 + d1)>
 // CHECK-LABEL: fold_static_stride_subview_with_affine_load_store_expand_shape_when_access_index_is_an_expression
 // CHECK-SAME: (%[[ARG0:.*]]: memref<1024x1024xf32>, %[[ARG1:.*]]: memref<1xf32>, %[[ARG2:.*]]: index)
 func.func @fold_static_stride_subview_with_affine_load_store_expand_shape_when_access_index_is_an_expression(%arg0: memref<1024x1024xf32>, %arg1: memref<1xf32>, %arg2: index) -> f32 {
@@ -578,14 +578,14 @@ func.func @fold_static_stride_subview_with_affine_load_store_expand_shape_when_a
 // CHECK-NEXT:  affine.for %[[ARG4:.*]] = 0 to 1024 {
 // CHECK-NEXT:   affine.for %[[ARG5:.*]] = 0 to 1020 {
 // CHECK-NEXT:    affine.for %[[ARG6:.*]] = 0 to 1 {
-// CHECK-NEXT:      %[[TMP1:.*]] = affine.apply #[[$MAP0]](%[[ARG3]], %[[ARG4]])[%[[ARG3]]]
-// CHECK-NEXT:      %[[TMP3:.*]] = affine.apply #[[$MAP1]]()[%[[ARG5]], %[[ARG6]]]
+// CHECK-NEXT:      %[[TMP1:.*]] = affine.apply #[[$MAP0]](%[[ARG3]], %[[ARG4]])
+// CHECK-NEXT:      %[[TMP3:.*]] = affine.apply #[[$MAP1]](%[[ARG5]], %[[ARG6]])
 // CHECK-NEXT:      affine.load %[[ARG0]][%[[TMP1]], %[[TMP3]]] : memref<1024x1024xf32>
 
 // -----
 
-// CHECK-DAG: #[[$MAP0:.*]] = affine_map<()[s0] -> (s0 * 1024)>
-// CHECK-DAG: #[[$MAP1:.*]] = affine_map<()[s0, s1] -> (s0 + s1)>
+// CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0) -> (d0 * 1024)>
+// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d0 + d1)>
 // CHECK-LABEL: fold_static_stride_subview_with_affine_load_store_expand_shape_with_constant_access_index
 // CHECK-SAME: (%[[ARG0:.*]]: memref<1024x1024xf32>, %[[ARG1:.*]]: memref<1xf32>, %[[ARG2:.*]]: index)
 func.func @fold_static_stride_subview_with_affine_load_store_expand_shape_with_constant_access_index(%arg0: memref<1024x1024xf32>, %arg1: memref<1xf32>, %arg2: index) -> f32 {
@@ -608,8 +608,8 @@ func.func @fold_static_stride_subview_with_affine_load_store_expand_shape_with_c
 // CHECK-NEXT:   affine.for %[[ARG4:.*]] = 0 to 1024 {
 // CHECK-NEXT:    affine.for %[[ARG5:.*]] = 0 to 1020 {
 // CHECK-NEXT:     affine.for %[[ARG6:.*]] = 0 to 1 {
-// CHECK-NEXT:      %[[TMP1:.*]] = affine.apply #[[$MAP0]]()[%[[ARG3]]]
-// CHECK-NEXT:      %[[TMP2:.*]] = affine.apply #[[$MAP1]]()[%[[ARG5]], %[[ARG6]]]
+// CHECK-NEXT:      %[[TMP1:.*]] = affine.apply #[[$MAP0]](%[[ARG3]])
+// CHECK-NEXT:      %[[TMP2:.*]] = affine.apply #[[$MAP1]](%[[ARG5]], %[[ARG6]])
 // CHECK-NEXT:      memref.load %[[ARG0]][%[[TMP1]], %[[TMP2]]] : memref<1024x1024xf32>
 
 // -----
@@ -678,7 +678,7 @@ func.func @fold_load_keep_nontemporal(%arg0 : memref<12x32xf32>, %arg1 : index,
 // -----
 
 // CHECK-LABEL: func @fold_store_keep_nontemporal(
-//      CHECK:   memref.store %{{.+}}, %{{.+}}[%{{.+}}, %{{.+}}]  {nontemporal = true} : memref<12x32xf32> 
+//      CHECK:   memref.store %{{.+}}, %{{.+}}[%{{.+}}, %{{.+}}]  {nontemporal = true} : memref<12x32xf32>
 func.func @fold_store_keep_nontemporal(%arg0 : memref<12x32xf32>, %arg1 : index, %arg2 : index, %arg3 : index, %arg4 : index, %arg5 : f32) {
   %0 = memref.subview %arg0[%arg1, %arg2][4, 4][2, 3] :
     memref<12x32xf32> to memref<4x4xf32, strided<[64, 3], offset: ?>>