[mlir][affine] Make [de]linearize_index a valid source of dims (#138929)

There's a sense in which affine.linearize_index and
affine.delinearize_index are special-cases of affine.apply (which get
their own ops to enable better code generation and more accurate
canonicalization). Therefore, allow these operations to be dimension
operands for operations like affine.load just like affine.apply can be.

Author: krzysz00

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

@@ -305,6 +305,8 @@ bool mlir::affine::isValidDim(Value value) {
 // *) It is valid as a symbol.
 // *) It is an induction variable.
 // *) It is the result of an affine apply operation with dimension id operands.
+// *) It is the result of a more specialized index transformation (ex.
+// delinearize_index or linearize_index) with dimension id operands.
 bool mlir::affine::isValidDim(Value value, Region *region) {
   // The value must be an index type.
   if (!value.getType().isIndex())
@@ -325,6 +327,11 @@ bool mlir::affine::isValidDim(Value value, Region *region) {
   // Affine apply operation is ok if all of its operands are ok.
   if (auto applyOp = dyn_cast<AffineApplyOp>(op))
     return applyOp.isValidDim(region);
+  // delinearize_index and linearize_index are special forms of apply
+  // and so are valid dimensions if all their arguments are valid dimensions.
+  if (isa<AffineDelinearizeIndexOp, AffineLinearizeIndexOp>(op))
+    return llvm::all_of(op->getOperands(),
+                        [&](Value arg) { return ::isValidDim(arg, region); });
   // The dim op is okay if its operand memref/tensor is defined at the top
   // level.
   if (auto dimOp = dyn_cast<ShapedDimOpInterface>(op))

mlir/test/Dialect/Affine/invalid.mlir

@@ -544,6 +544,34 @@ func.func @dynamic_dimension_index() {
 
 // -----
 
+func.func @dynamic_linearized_index() {
+  "unknown.region"() ({
+    %idx = "unknown.test"() : () -> (index)
+    %memref = "unknown.test"() : () -> memref<?xf32>
+    %pos = affine.linearize_index [%idx, %idx] by (8) : index
+    // expected-error@below {{op operand cannot be used as a dimension id}}
+    affine.load %memref[%pos] : memref<?xf32>
+    "unknown.terminator"() : () -> ()
+  }) : () -> ()
+  return
+}
+
+// -----
+
+func.func @dynamic_delinearized_index() {
+  "unknown.region"() ({
+    %idx = "unknown.test"() : () -> (index)
+    %memref = "unknown.test"() : () -> memref<?x?xf32>
+    %pos0, %pos1 = affine.delinearize_index %idx into (8) : index, index
+    // expected-error@below {{op operand cannot be used as a dimension id}}
+    affine.load %memref[%pos0, %pos1] : memref<?x?xf32>
+    "unknown.terminator"() : () -> ()
+  }) : () -> ()
+  return
+}
+
+// -----
+
 #map = affine_map<() -> ()>
 #map1 = affine_map<() -> (1)>
 func.func @no_lower_bound() {

mlir/test/Dialect/Affine/ops.mlir

@@ -148,6 +148,23 @@ func.func @valid_symbol_affine_scope(%n : index, %A : memref<?xf32>) {
 
 // -----
 
+// Test dimension constraints for linearize_index and delinearize_index
+
+// CHECK-LABEL: func @valid_dim_linearize_delinearize
+func.func @valid_dim_linearize_delinearize(%m : index, %n : index, %A : memref<?xf32>, %B: memref<?x32x?xf32>) {
+    affine.for %0 = 0 to %m {
+      affine.for %1 = 0 to %n {
+        %load_idx = affine.linearize_index disjoint [%0, %1] by (%m, %n) : index
+        %store_idx0, %store_idx1 = affine.delinearize_index %n into (32) : index, index
+        %v = affine.load %A[%load_idx] : memref<?xf32>
+        affine.store %v, %B[%0, %store_idx1, %store_idx0] : memref<?x32x?xf32>
+      }
+    }
+  return
+}
+
+// -----
+
 // Test the fact that module op always provides an affine scope.
 
 %idx = "test.foo"() : () -> (index)
@@ -309,7 +326,7 @@ func.func @linearize_mixed(%index0: index, %index1: index, %index2: index, %basi
 module {
   func.func @gpu_launch_affine() {
     %c1 = arith.constant 1 : index
-    gpu.launch blocks(%arg0, %arg1, %arg2) in (%arg6 = %c1, %arg7 = %c1, %arg8 = %c1) 
+    gpu.launch blocks(%arg0, %arg1, %arg2) in (%arg6 = %c1, %arg7 = %c1, %arg8 = %c1)
     threads(%arg3, %arg4, %arg5) in (%arg9 = %c1, %arg10 = %c1, %arg11 = %c1) {
       %thread_id_x = gpu.thread_id  x
       %c128 = arith.constant 128 : index