[Mlir-commits] [mlir] [MLIR][Affine] Fix affine.apply verifier and add functionality to demote invalid symbols to dims (PR #128289)
Uday Bondhugula
llvmlistbot at llvm.org
Thu Mar 27 02:59:33 PDT 2025
https://github.com/bondhugula updated https://github.com/llvm/llvm-project/pull/128289
>From 5bb29d2e39f5c6514a943714e71316665b62c344 Mon Sep 17 00:00:00 2001
From: Uday Bondhugula <uday at polymagelabs.com>
Date: Sun, 16 Feb 2025 10:41:10 +0530
Subject: [PATCH] [MLIR][Affine] Fix affine.apply verifier and add
functionality to demote invalid symbols to dims
Fixes: https://github.com/llvm/llvm-project/issues/120189,
https://github.com/llvm/llvm-project/issues/128403
Fix affine.apply verifier to reject symbolic operands that are valid
dims for affine purposes. This doesn't affect other users in other
context 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 PR also fixes test cases where dimensional positions should have
been ideally used with affine.apply (for affine loop IVs for example).
---
mlir/lib/Dialect/Affine/IR/AffineOps.cpp | 63 ++++++++++++++++++-
mlir/test/Dialect/Affine/canonicalize.mlir | 4 +-
mlir/test/Dialect/Affine/invalid.mlir | 14 +++++
.../Dialect/MemRef/fold-memref-alias-ops.mlir | 36 +++++------
4 files changed, 95 insertions(+), 22 deletions(-)
diff --git a/mlir/lib/Dialect/Affine/IR/AffineOps.cpp b/mlir/lib/Dialect/Affine/IR/AffineOps.cpp
index 9c5b9e82cd5e0..3644938092231 100644
--- a/mlir/lib/Dialect/Affine/IR/AffineOps.cpp
+++ b/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 +1368,62 @@ 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.
+// This function canonicalizes symbols 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;
+
+ assert(mapOrSet->getNumInputs() == operands.size() &&
+ "map/set inputs must match number of operands");
+
+ auto *context = mapOrSet->getContext();
+ SmallVector<Value, 8> resultOperands;
+ resultOperands.reserve(operands.size());
+ SmallVector<Value, 8> remappedDims;
+ remappedDims.reserve(operands.size());
+ unsigned nextSym = 0;
+ unsigned nextDim = 0;
+ unsigned oldNumDims = mapOrSet->getNumDims();
+ SmallVector<AffineExpr, 8> symRemapping(mapOrSet->getNumSymbols());
+ for (unsigned i = 0, e = mapOrSet->getNumInputs(); i != e; ++i) {
+ if (i >= oldNumDims) {
+ 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);
+ resultOperands.push_back(operands[i]);
+ }
+ } else {
+ resultOperands.push_back(operands[i]);
+ }
+ }
+
+ resultOperands.insert(resultOperands.begin() + oldNumDims,
+ remappedDims.begin(), remappedDims.end());
+ 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 +1438,7 @@ 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());
diff --git a/mlir/test/Dialect/Affine/canonicalize.mlir b/mlir/test/Dialect/Affine/canonicalize.mlir
index d39c0c6e41df2..e56079c1cccd4 100644
--- a/mlir/test/Dialect/Affine/canonicalize.mlir
+++ b/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
diff --git a/mlir/test/Dialect/Affine/invalid.mlir b/mlir/test/Dialect/Affine/invalid.mlir
index 9bbd19c381163..9703c05fff8f6 100644
--- a/mlir/test/Dialect/Affine/invalid.mlir
+++ b/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 at above {{dimensional operand cannot be used as a symbol}}
+ }
+ }
+ }
+ return
+}
diff --git a/mlir/test/Dialect/MemRef/fold-memref-alias-ops.mlir b/mlir/test/Dialect/MemRef/fold-memref-alias-ops.mlir
index 327cacf7d9a20..62f58d5585de4 100644
--- a/mlir/test/Dialect/MemRef/fold-memref-alias-ops.mlir
+++ b/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: ?>>
More information about the Mlir-commits
mailing list