[MLIR] Simplify Semi Affine Expressions Using flattening

[arnab-oss]

Summary

For the semi affine expressions, whenever rhs of a floordiv, ceildiv, mod
or product expression is a symbolic expression, we introduce a local variable
representing the result, and store the floordiv/ceildiv, mod or product
affine expression in LocalExprs. In this way the expression is flattened,
and trivial addition and subtraction related simplifications are performed.

Also rule based matching for detecting and transforming "expr - q * (expr floordiv q)"
to "expr mod q", where q is a symbolic exxpression, in simplifyAdd function.

[bondhugula]

Can you split the two commits out into separate revisions (with different titles)? The first one can be set as the parent revision of the second one.

[arnab-oss]

Remove the second commit and keep only the first commit doing simplification of semi affine map in this PR.

[springerm]

Thanks for the contribution, this is really useful. Some first comments, will do a deeper dive in the next days...

mlir/lib/IR/AffineExpr.cpp

932

It seems like you're using -1 in the code below. Comments need updating?

937

d0

937–938

Would be good to have a full example with coefficients, indices and indexToExprMap somewhere here.

967

nit: store in variable to avoid writing it 3 times

980

involving

982

space

984

Use llvm::enumerate, so that you don't have to maintain an idx variable.

986

Is there an idx++ missing here?

999

delete

1003–1004

Is it possible that there is already an entry at {lhsPos, -1}?

1021–1024

This piece of code is a common pattern. I would create a lambda function and reuse it.

Something like

auto addEntry = [&](std::pair<...> index, int64_t coefficient, AffineExpr expr) {
  assert that index is not yet in indices

  indices.push_back(index);
  coefficients.insert(...);
  indexToExprMap.insert(...)
};

utils/arcanist/clang-format.sh

54

I think this is not supposed to be part of the commit.

[arnab-oss]

Addressed all the comments by springerm.

[arnab-oss]

Added an example showing the contents of indices, indexToExprMap and coefficients.

[arnab-oss]

Removed unwanted changes from clang-format.sh

[bondhugula]

Thanks for this really useful simplification that catches most of the things needed in practice. Mostly minor comments below.

mlir/include/mlir/IR/AffineExpr.h

296–298

Nit: You can rewrap to use the whole width (80 columns). I know it was the existing comment that didn't.

mlir/include/mlir/IR/AffineExprVisitor.h

303

No comma before when.

312

maybe -> may be

312

Please rephrase the first sentence - it isn't really clear.

which represents -> representing ?

314

Spell fix.

315

in proper -> in the proper

mlir/lib/IR/AffineExpr.cpp

954

Function to add entries -> Adds entries

1063–1066

Code comments here.

1094

product, and the -> product; the ...

1133

Using semi-affine (which hyphen) consistently.

1232–1236

But result's size has not been set. If the expectation is for the caller to set it, there has to be an assertion here on the expected size of result.

1263

This should be:
SmallVectorImpl<int64_t> &lhs = ...
?

[dcaballe]

Thanks a lot for the contribution! LGTM. Just some minor comments.

mlir/lib/IR/AffineExpr.cpp

942

givem -> given

969

j++ -> ++j

mlir/test/Dialect/Affine/simplify-affine-structures.mlir

493

Please, add a CHECK-LABEL for this test

504

Could you please split the test into multiple ones (one for mod, one for floordiv, etc.)? It would help readability and be very helpful to quickly know which simplification is broken when the time comes.

510

Maybe I'm wrong but, unlike affine maps, I think the order of the instructions shouldn't change here so we could remove the -DAG part (?)

[arnab-oss]

Addressed all the comments by @bondhugula and @dcaballe.

[springerm]

mlir/lib/IR/AffineExpr.cpp

969

nit: I would split this in two loops, since you have the j < numDims check below.

1002

I still think there's an idx++ missing here. Can you explain the reasoning for this?

[arnab-oss]

Fixed a bug in the loop iterating over localExprs vector. Addressed comments
by @springerm.

[bondhugula]

In D112808#3118161, @arnab-oss wrote:

Fixed a bug in the loop iterating over localExprs vector. Addressed comments
by @springerm.

Could you add a test case that exercises that part of the code? (if not already)

[arnab-oss]

Added an example which is giving correct output now after the previous update bug fix.

[arnab-oss]

In D112808#3122209, @bondhugula wrote:

In D112808#3118161, @arnab-oss wrote:

Fixed a bug in the loop iterating over localExprs vector. Addressed comments
by @springerm.

Could you add a test case that exercises that part of the code? (if not already)

Added. Thanks.

mlir/lib/IR/AffineExpr.cpp

984

I cannot use llvm::enumerate and replace idx with iterator.index() as idx won't be increasing in every iteration.

986

No

1002

Oh you are correct. In fact this change made me uncover another bug in the code.

1003–1004

No

1263

If I do this, I cannot directly use lhs to initialise dividend later in the code.

mlir/test/Dialect/Affine/simplify-affine-structures.mlir

493

Using CHECK-LABEL for function name and CHECK-SAME for capturing arhguments leads to test failure.

It shows error message "PRODUCT" not defined which I had defined earlier while capturing the map using CHECK-DAG.

[dcaballe]

mlir/test/Dialect/Affine/simplify-affine-structures.mlir

493

You have to use $ to capture variables before LABEL. Something like: #[[$MOD:.*]] = affine_map...

[arnab-oss]

Added CHECK-LABEL in unit tests.

[bondhugula]

LGTM - thanks very much!

[arnab-oss]

ping @springerm and @dcaballe , can you please let me know whether the revision is ok or if any more changes are required to be made? Thanks in advance!

mlir/test/Dialect/Affine/simplify-affine-structures.mlir

493

Done.

[dcaballe]

In D112808#3129562, @arnab-oss wrote:

ping @springerm and @dcaballe , can you please let me know whether the revision is ok or if any more changes are required to be made? Thanks in advance!

Thanks for addressing the feedback! The tests look good to me. I'll leave the final approval to @springerm.

[bondhugula]

There were three whitespace errors in the test case and the commit title capitalization that I fixed before committing.

mlir/include/mlir/IR/AffineExpr.h

U AffineExpr::dyn_cast_or_null() const {
   return (!*this || !isa<U>()) ? U(nullptr) : U(expr);
 }
 template <typename U>
 U AffineExpr::cast() const {
   assert(isa<U>());
   return U(expr);
 }
 
-/// Simplify an affine expression by flattening and some amount of
-/// simple analysis. This has complexity linear in the number of nodes in
+/// Simplify an affine expression by flattening and some amount
+/// of simple analysis. This has complexity linear in the number of nodes in
 /// 'expr'. Returns the simplified expression, which is the same as the input
-///  expression if it can't be simplified.
+/// expression if it can't be simplified. When `expr` is semi-affine, a
+/// simplified semi-affine expression is constructed in the sorted order of
+/// dimension and symbol positions.
 AffineExpr simplifyAffineExpr(AffineExpr expr, unsigned numDims,
                               unsigned numSymbols);
 
 namespace detail {
 template <int N>
 void bindDims(MLIRContext *ctx) {}
 
 template <int N, typename AffineExprTy, typename... AffineExprTy2>

mlir/include/mlir/IR/AffineExprVisitor.h

protected:
   // Add a local identifier (needed to flatten a mod, floordiv, ceildiv expr).
   // The local identifier added is always a floordiv of a pure add/mul affine
   // function of other identifiers, coefficients of which are specified in
   // dividend and with respect to a positive constant divisor. localExpr is the
   // simplified tree expression (AffineExpr) corresponding to the quantifier.
   virtual void addLocalFloorDivId(ArrayRef<int64_t> dividend, int64_t divisor,
                                   AffineExpr localExpr);
 
+  /// Add a local identifier (needed to flatten a mod, floordiv, ceildiv, mul
+  /// expr), when the rhs is a symbolic expression. The local identifier added
+  /// may be a floordiv, ceildiv, mul or mod of a pure affine/semi-affine
+  /// function of other identifiers, coefficients of which are specified in the
+  /// lhs of the mod, floordiv, ceildiv or mul expression and with respect to a
+  /// symbolic rhs expression. `localExpr` is the simplified tree expression
+  /// (AffineExpr) corresponding to the quantifier.
+  virtual void addLocalIdSemiAffine(AffineExpr localExpr);
+
 private:
+  /// Adds `expr`, which maybe mod, ceildiv, floordiv or mod expression, which
+  /// represents the affine expression corresponding to the quantifier
+  /// introduced as the local variable corrsponding to `expr`. If the quantifier
+  /// is already present, we put the coefficient in proper index of `result`,
+  /// otherwise we add a new local variable and put the coefficient there.
+  void addLocalVariableSemiAffine(AffineExpr expr,
+                                  SmallVectorImpl<int64_t> &result);
+
   // t = expr floordiv c   <=> t = q, c * q <= expr <= c * q + c - 1
   // A floordiv is thus flattened by introducing a new local variable q, and
   // replacing that expression with 'q' while adding the constraints
   // c * q <= expr <= c * q + c - 1 to localVarCst (done by
   // FlatAffineConstraints::addLocalFloorDiv).
   //
   // A ceildiv is similarly flattened:
   // t = expr ceildiv c   <=> t =  (expr + c - 1) floordiv c

mlir/lib/IR/AffineExpr.cpp

AffineExpr mlir::getAffineExprFromFlatForm(ArrayRef<int64_t> flatExprs,
 
   // Constant term.
   int64_t constTerm = flatExprs[flatExprs.size() - 1];
   if (constTerm != 0)
     expr = expr + constTerm;
   return expr;
 }
 
+/// Constructs a semi-affine expression from a flat ArrayRef. If there are
+/// local identifiers (neither dimensional nor symbolic) that appear in the sum
+/// of products expression, `localExprs` is expected to have the AffineExprs for
+/// it, and is substituted into. The ArrayRef `flatExprs` is expected to be in
+/// the format [dims, symbols, locals, constant term]. The semi-affine
+/// expression is constructed in the sorted order of dimension and symbol
+/// position numbers. Note:  local expressions/ids are used for mod, div as well
+/// as symbolic RHS terms for terms that are not pure affine.
+static AffineExpr getSemiAffineExprFromFlatForm(ArrayRef<int64_t> flatExprs,
+                                                unsigned numDims,
+                                                unsigned numSymbols,
+                                                ArrayRef<AffineExpr> localExprs,
+                                                MLIRContext *context) {
+  assert(!flatExprs.empty() && "flatExprs cannot be empty");
+
+  // Assert expected numLocals = flatExprs.size() - numDims - numSymbols - 1.
+  assert(flatExprs.size() - numDims - numSymbols - 1 == localExprs.size() &&
+         "unexpected number of local expressions");
+
+  AffineExpr expr = getAffineConstantExpr(0, context);
+
+  // We design indices as a pair which help us present the semi-affine map as
+  // sum of product where terms are sorted based on dimension or symbol
+  // position: <keyA, keyB> for expressions of the form dimension * symbol,
+  // where keyA is the position number of the dimension and keyB is the
+  // position number of the symbol. For dimensional expressions we set the index
+  // as (position number of the dimension, -1), as we want dimensional
+  // expressions to appear before symbolic and product of dimensional and
+  // symbolic expressions having the dimension with the same position number.
+  // For symbolic expression set the index as (position number of the symbol,
+  // maximum of last dimension and symbol position) number. For example, we want
+  // the expression we are constructing to look something like: d0 + d0 * s0 +
+  // s0 + d1*s1 + s1.
+
+  // Stores the affine expression corresponding to a given index.
+  DenseMap<std::pair<unsigned, signed>, AffineExpr> indexToExprMap;
+  // Stores the constant coefficient value corresponding to a givem
+  // dimension, symbol or a non-pure affine expression stored in `localExprs`.
+  DenseMap<std::pair<unsigned, signed>, int64_t> coefficients;
+  // Stores the indices as defined above, and later sorted to produce
+  // the semi-affine expression in the desired form.
+  SmallVector<std::pair<unsigned, signed>, 8> indices;
+
+  // Example: expression = d0 + d0 * s0 + 2 * s0.
+  // indices = [{0,-1}, {0, 0}, {0, 1}]
+  // coefficients = [{{0, -1}, 1}, {{0, 0}, 1}, {{0, 1}, 2}]
+  // indexToExprMap = [{{0, -1}, d0}, {{0, 0}, d0 * s0}, {{0, 1}, s0}]
+
+  // Function to add entries to `indexToExprMap`, `coefficients` and `indices`.
+  auto addEntry = [&](std::pair<unsigned, signed> index, int64_t coefficient,
+                      AffineExpr expr) {
+    assert(std::find(indices.begin(), indices.end(), index) == indices.end() &&
+           "Key is already present in indices vector and overwriting will "
+           "happen in `indexToExprMap` and `coefficients`!");
+
+    indices.push_back(index);
+    coefficients.insert({index, coefficient});
+    indexToExprMap.insert({index, expr});
+  };
+
+  // Design indices for dimensional or symbolic terms, and store the indices,
+  // constant coefficient corresponding to the indices in `coefficients` map,
+  // and affine expression corresponding to indices in `indexToExprMap` map.
+  for (unsigned j = 0; j < numDims + numSymbols; j++) {
+    if (flatExprs[j] == 0)
+      continue;
+    if (j < numDims) {
+      // For dimensional expressions we set the index as <position number of the
+      // dimension, 0>, as we want dimensional expressions to appear before
+      // symbolic ones and products of dimensional and symbolic expressions
+      // having the dimension with the same position number.
+      std::pair<unsigned, signed> indexEntry(j, -1);
+      addEntry(indexEntry, flatExprs[j], getAffineDimExpr(j, context));
+    } else {
+      // For symbolic expression set the index as <position number
+      // of the symbol, max(dimCount, symCount)> number,
+      // as we want symbolic expressions with the same positional number to
+      // appear after dimensional expressions having the same positional number.
+      std::pair<unsigned, signed> indexEntry(j - numDims,
+                                             std::max(numDims, numSymbols));
+      addEntry(indexEntry, flatExprs[j],
+               getAffineSymbolExpr(j - numDims, context));
+    }
+  }
+  // Denotes semi-affine product, modulo or division terms, which has been added
+  // to the `indexToExpr` map.
+  SmallVector<bool, 4> addedToMap(flatExprs.size() - numDims - numSymbols - 1,
+                                  false);
+  unsigned idx = 0;
+  unsigned lhsPos, rhsPos;
+  // Construct indices for product terms involving dimension, symbol or constant
+  // as lhs/rhs, and store the indices, constant coefficient corresponding to
+  // the indices in `coefficients` map, and affine expression corresponding to
+  // in indices in `indexToExprMap` map.
+  for (AffineExpr expr : localExprs) {
+    if (flatExprs[numDims + numSymbols + idx] == 0)
+      continue;
+    AffineExpr lhs = expr.cast<AffineBinaryOpExpr>().getLHS();
+    AffineExpr rhs = expr.cast<AffineBinaryOpExpr>().getRHS();
+    if (!((lhs.isa<AffineDimExpr>() || lhs.isa<AffineSymbolExpr>()) &&
+          (rhs.isa<AffineDimExpr>() || rhs.isa<AffineSymbolExpr>() ||
+           rhs.isa<AffineConstantExpr>()))) {
+      idx++;
+      continue;
+    }
+    if (rhs.isa<AffineConstantExpr>()) {
+      // For product/modulo/division expressions, when rhs of modulo/division
+      // expression is constant, we put 0 in place of keyB, because we want
+      // them to appear earlier in the semi-affine expression we are
+      // constructing. When rhs is constant, we place 0 in place of keyB.
+      if (lhs.isa<AffineDimExpr>()) {
+        lhsPos = lhs.cast<AffineDimExpr>().getPosition();
+        std::pair<unsigned, signed> indexEntry(lhsPos, -1);
+        addEntry(indexEntry, flatExprs[numDims + numSymbols + idx], expr);
+      } else {
+        lhsPos = lhs.cast<AffineSymbolExpr>().getPosition();
+        std::pair<unsigned, signed> indexEntry(lhsPos,
+                                               std::max(numDims, numSymbols));
+        addEntry(indexEntry, flatExprs[numDims + numSymbols + idx], expr);
+      }
+    } else if (lhs.isa<AffineDimExpr>()) {
+      // For product/modulo/division expressions having lhs as dimension and rhs
+      // as symbol, we order the terms in the semi-affine expression based on
+      // the pair: <keyA, keyB> for expressions of the form dimension * symbol,
+      // where keyA is the position number of the dimension and keyB is the
+      // position number of the symbol.
+      lhsPos = lhs.cast<AffineDimExpr>().getPosition();
+      rhsPos = rhs.cast<AffineSymbolExpr>().getPosition();
+      std::pair<unsigned, signed> indexEntry(lhsPos, rhsPos);
+      addEntry(indexEntry, flatExprs[numDims + numSymbols + idx], expr);
+    } else {
+      // For product/modulo/division expressions having both lhs and rhs as
+      // symbol, we design indices as a pair: <keyA, keyB> for expressions
+      // of the form dimension * symbol, where keyA is the position number of
+      // the dimension and keyB is the position number of the symbol.
+      lhsPos = lhs.cast<AffineSymbolExpr>().getPosition();
+      rhsPos = rhs.cast<AffineSymbolExpr>().getPosition();
+      std::pair<unsigned, signed> indexEntry(lhsPos, rhsPos);
+      addEntry(indexEntry, flatExprs[numDims + numSymbols + idx], expr);
+    }
+    addedToMap[idx] = true;
+    idx++;
+  }
+
+  // Constructing the simplified semi-affine sum of product/division/mod
+  // expression from the flattened form in the desired sorted order of indices
+  // of the various individual product/division/mod expressions.
+  std::sort(indices.begin(), indices.end());
+  for (const std::pair<unsigned, unsigned> index : indices) {
+    assert(indexToExprMap.lookup(index) &&
+           "cannot find key in `indexToExprMap` map");
+    expr = expr + indexToExprMap.lookup(index) * coefficients.lookup(index);
+  }
+
+  // Local identifiers.
+  for (unsigned j = numDims + numSymbols, e = flatExprs.size() - 1; j < e;
+       j++) {
+    if (flatExprs[j] == 0 || addedToMap[j - numDims - numSymbols])
+      continue;
+    auto term = localExprs[j - numDims - numSymbols] * flatExprs[j];
+    expr = expr + term;
+  }
+
+  // Constant term.
+  int64_t constTerm = flatExprs.back();
+  if (constTerm != 0)
+    expr = expr + constTerm;
+  return expr;
+}
+
 SimpleAffineExprFlattener::SimpleAffineExprFlattener(unsigned numDims,
                                                      unsigned numSymbols)
     : numDims(numDims), numSymbols(numSymbols), numLocals(0) {
   operandExprStack.reserve(8);
 }
 
+// In pure affine t = expr * c, we multiply each coefficient of lhs with c.
+//
+// In case of semi affine multiplication expressions, t = expr * symbolic_expr,
+// introduce a local variable p (= expr * symbolic_expr), and the affine
+// expression expr * symbolic_expr is added to `localExprs`.
 void SimpleAffineExprFlattener::visitMulExpr(AffineBinaryOpExpr expr) {
   assert(operandExprStack.size() >= 2);
-  // This is a pure affine expr; the RHS will be a constant.
-  assert(expr.getRHS().isa<AffineConstantExpr>());
-  // Get the RHS constant.
-  auto rhsConst = operandExprStack.back()[getConstantIndex()];
+  SmallVector<int64_t, 8> rhs = operandExprStack.back();
   operandExprStack.pop_back();
-  // Update the LHS in place instead of pop and push.
-  auto &lhs = operandExprStack.back();
+  SmallVector<int64_t, 8> &lhs = operandExprStack.back();
+
+  // Flatten semi-affine multiplication expressions by introducing a local
+  // variable in place of the product, and the affine expression
+  // corresponding to the quantifier is added to `localExprs`.
+  if (!expr.getRHS().isa<AffineConstantExpr>()) {
+    MLIRContext *context = expr.getContext();
+    AffineExpr a = getAffineExprFromFlatForm(lhs, numDims, numSymbols,
+                                             localExprs, context);
+    AffineExpr b = getAffineExprFromFlatForm(rhs, numDims, numSymbols,
+                                             localExprs, context);
+    addLocalVariableSemiAffine(a * b, lhs);
+    return;
+  }
+
+  // Get the RHS constant.
+  auto rhsConst = rhs[getConstantIndex()];
   for (unsigned i = 0, e = lhs.size(); i < e; i++) {
     lhs[i] *= rhsConst;
   }
 }
 
 void SimpleAffineExprFlattener::visitAddExpr(AffineBinaryOpExpr expr) {
   assert(operandExprStack.size() >= 2);
   const auto &rhs = operandExprStack.back();
   auto &lhs = operandExprStack[operandExprStack.size() - 2];
   assert(lhs.size() == rhs.size());
   // Update the LHS in place.
   for (unsigned i = 0, e = rhs.size(); i < e; i++) {
     lhs[i] += rhs[i];
   }
   // Pop off the RHS.
   operandExprStack.pop_back();
 }
 
 //
 // t = expr mod c   <=>  t = expr - c*q and c*q <= expr <= c*q + c - 1
 //
 // A mod expression "expr mod c" is thus flattened by introducing a new local
 // variable q (= expr floordiv c), such that expr mod c is replaced with
 // 'expr - c * q' and c * q <= expr <= c * q + c - 1 are added to localVarCst.
+//
+// In case of semi affine modulo expressions, t = expr mod symbolic_expr,
+// introduce a local variable m (= expr mod symbolic_expr), and the affine
+// expression expr mod symbolic_expr is added to `localExprs`.
 void SimpleAffineExprFlattener::visitModExpr(AffineBinaryOpExpr expr) {
   assert(operandExprStack.size() >= 2);
-  // This is a pure affine expr; the RHS will be a constant.
-  assert(expr.getRHS().isa<AffineConstantExpr>());
-  auto rhsConst = operandExprStack.back()[getConstantIndex()];
+
+  SmallVector<int64_t, 8> rhs = operandExprStack.back();
   operandExprStack.pop_back();
-  auto &lhs = operandExprStack.back();
+  SmallVector<int64_t, 8> &lhs = operandExprStack.back();
+  MLIRContext *context = expr.getContext();
+
+  // Flatten semi affine modulo expressions by introducing a local
+  // variable in place of the modulo value, and the affine expression
+  // corresponding to the quantifier is added to `localExprs`.
+  if (!expr.getRHS().isa<AffineConstantExpr>()) {
+    AffineExpr dividendExpr = getAffineExprFromFlatForm(
+        lhs, numDims, numSymbols, localExprs, context);
+    AffineExpr divisorExpr = getAffineExprFromFlatForm(rhs, numDims, numSymbols,
+                                                       localExprs, context);
+    AffineExpr modExpr = dividendExpr % divisorExpr;
+    addLocalVariableSemiAffine(modExpr, lhs);
+    return;
+  }
+
+  int64_t rhsConst = rhs[getConstantIndex()];
   // TODO: handle modulo by zero case when this issue is fixed
   // at the other places in the IR.
   assert(rhsConst > 0 && "RHS constant has to be positive");
 
   // Check if the LHS expression is a multiple of modulo factor.
   unsigned i, e;
   for (i = 0, e = lhs.size(); i < e; i++)
     if (lhs[i] % rhsConst != 0)
void SimpleAffineExprFlattener::visitModExpr(AffineBinaryOpExpr expr) {
   // Simplify the numerator and the denominator.
   if (gcd != 1) {
     for (unsigned i = 0, e = floorDividend.size(); i < e; i++)
       floorDividend[i] = floorDividend[i] / static_cast<int64_t>(gcd);
   }
   int64_t floorDivisor = rhsConst / static_cast<int64_t>(gcd);
 
   // Construct the AffineExpr form of the floordiv to store in localExprs.
-  MLIRContext *context = expr.getContext();
-  auto dividendExpr = getAffineExprFromFlatForm(
+
+  AffineExpr dividendExpr = getAffineExprFromFlatForm(
       floorDividend, numDims, numSymbols, localExprs, context);
-  auto divisorExpr = getAffineConstantExpr(floorDivisor, context);
-  auto floorDivExpr = dividendExpr.floorDiv(divisorExpr);
+  AffineExpr divisorExpr = getAffineConstantExpr(floorDivisor, context);
+  AffineExpr floorDivExpr = dividendExpr.floorDiv(divisorExpr);
   int loc;
   if ((loc = findLocalId(floorDivExpr)) == -1) {
     addLocalFloorDivId(floorDividend, floorDivisor, floorDivExpr);
     // Set result at top of stack to "lhs - rhsConst * q".
     lhs[getLocalVarStartIndex() + numLocals - 1] = -rhsConst;
   } else {
     // Reuse the existing local id.
     lhs[getLocalVarStartIndex() + loc] = -rhsConst;

 }
 
 void SimpleAffineExprFlattener::visitConstantExpr(AffineConstantExpr expr) {
   operandExprStack.emplace_back(SmallVector<int64_t, 32>(getNumCols(), 0));
   auto &eq = operandExprStack.back();
   eq[getConstantIndex()] = expr.getValue();
 }
 
+void SimpleAffineExprFlattener::addLocalVariableSemiAffine(
+    AffineExpr expr, SmallVectorImpl<int64_t> &result) {
+  int loc;
+  if ((loc = findLocalId(expr)) == -1)
+    addLocalIdSemiAffine(expr);
+  std::fill(result.begin(), result.end(), 0);
+  if (loc == -1)
+    result[getLocalVarStartIndex() + numLocals - 1] = 1;
+  else
+    result[getLocalVarStartIndex() + loc] = 1;
+}
+
 // t = expr floordiv c   <=> t = q, c * q <= expr <= c * q + c - 1
 // A floordiv is thus flattened by introducing a new local variable q, and
 // replacing that expression with 'q' while adding the constraints
 // c * q <= expr <= c * q + c - 1 to localVarCst (done by
 // FlatAffineConstraints::addLocalFloorDiv).
 //
 // A ceildiv is similarly flattened:
 // t = expr ceildiv c   <=> t =  (expr + c - 1) floordiv c
+//
+// In case of semi affine division expressions, t = expr floordiv symbolic_expr
+// or t = expr ceildiv symbolic_expr, introduce a local variable q (= expr
+// floordiv/ceildiv symbolic_expr), and the affine floordiv/ceildiv is added to
+// `localExprs`.
 void SimpleAffineExprFlattener::visitDivExpr(AffineBinaryOpExpr expr,
                                              bool isCeil) {
   assert(operandExprStack.size() >= 2);
-  assert(expr.getRHS().isa<AffineConstantExpr>());
+
+  MLIRContext *context = expr.getContext();
+  SmallVector<int64_t, 8> rhs = operandExprStack.back();
+  operandExprStack.pop_back();
+  SmallVector<int64_t, 8> &lhs = operandExprStack.back();
+
+  // Flatten semi affine division expressions by introducing a local
+  // variable in place of the quotient, and the affine expression corresponding
+  // to the quantifier is added to `localExprs`.
+  if (!expr.getRHS().isa<AffineConstantExpr>()) {
+    AffineExpr a = getAffineExprFromFlatForm(lhs, numDims, numSymbols,
+                                             localExprs, context);
+    AffineExpr b = getAffineExprFromFlatForm(rhs, numDims, numSymbols,
+                                             localExprs, context);
+    AffineExpr divExpr = isCeil ? a.ceilDiv(b) : a.floorDiv(b);
+    addLocalVariableSemiAffine(divExpr, lhs);
+    return;
+  }
 
   // This is a pure affine expr; the RHS is a positive constant.
-  int64_t rhsConst = operandExprStack.back()[getConstantIndex()];
+  int64_t rhsConst = rhs[getConstantIndex()];
   // TODO: handle division by zero at the same time the issue is
   // fixed at other places.
   assert(rhsConst > 0 && "RHS constant has to be positive");
-  operandExprStack.pop_back();
-  auto &lhs = operandExprStack.back();
 
   // Simplify the floordiv, ceildiv if possible by canceling out the greatest
   // common divisors of the numerator and denominator.
   uint64_t gcd = std::abs(rhsConst);
   for (unsigned i = 0, e = lhs.size(); i < e; i++)
     gcd = llvm::GreatestCommonDivisor64(gcd, std::abs(lhs[i]));
   // Simplify the numerator and the denominator.
   if (gcd != 1) {
     for (unsigned i = 0, e = lhs.size(); i < e; i++)
       lhs[i] = lhs[i] / static_cast<int64_t>(gcd);
   }
   int64_t divisor = rhsConst / static_cast<int64_t>(gcd);
   // If the divisor becomes 1, the updated LHS is the result. (The
   // divisor can't be negative since rhsConst is positive).
   if (divisor == 1)
     return;
 
   // If the divisor cannot be simplified to one, we will have to retain
   // the ceil/floor expr (simplified up until here). Add an existential
   // quantifier to express its result, i.e., expr1 div expr2 is replaced
   // by a new identifier, q.
-  MLIRContext *context = expr.getContext();
-  auto a =
+  AffineExpr a =
       getAffineExprFromFlatForm(lhs, numDims, numSymbols, localExprs, context);
-  auto b = getAffineConstantExpr(divisor, context);
+  AffineExpr b = getAffineConstantExpr(divisor, context);
 
   int loc;
-  auto divExpr = isCeil ? a.ceilDiv(b) : a.floorDiv(b);
+  AffineExpr divExpr = isCeil ? a.ceilDiv(b) : a.floorDiv(b);
   if ((loc = findLocalId(divExpr)) == -1) {
     if (!isCeil) {
       SmallVector<int64_t, 8> dividend(lhs);
       addLocalFloorDivId(dividend, divisor, divExpr);
     } else {
       // lhs ceildiv c <=>  (lhs + c - 1) floordiv c
       SmallVector<int64_t, 8> dividend(lhs);
       dividend.back() += divisor - 1;

 // The local identifier added is always a floordiv of a pure add/mul affine
 // function of other identifiers, coefficients of which are specified in
 // dividend and with respect to a positive constant divisor. localExpr is the
 // simplified tree expression (AffineExpr) corresponding to the quantifier.
 void SimpleAffineExprFlattener::addLocalFloorDivId(ArrayRef<int64_t> dividend,
                                                    int64_t divisor,
                                                    AffineExpr localExpr) {
   assert(divisor > 0 && "positive constant divisor expected");
-  for (auto &subExpr : operandExprStack)
+  for (SmallVector<int64_t, 8> &subExpr : operandExprStack)
     subExpr.insert(subExpr.begin() + getLocalVarStartIndex() + numLocals, 0);
   localExprs.push_back(localExpr);
   numLocals++;
   // dividend and divisor are not used here; an override of this method uses it.
 }
 
+void SimpleAffineExprFlattener::addLocalIdSemiAffine(AffineExpr localExpr) {
+  for (SmallVector<int64_t, 8> &subExpr : operandExprStack)
+    subExpr.insert(subExpr.begin() + getLocalVarStartIndex() + numLocals, 0);
+  localExprs.push_back(localExpr);
+  ++numLocals;
+}
+
 int SimpleAffineExprFlattener::findLocalId(AffineExpr localExpr) {
   SmallVectorImpl<AffineExpr>::iterator it;
   if ((it = llvm::find(localExprs, localExpr)) == localExprs.end())
     return -1;
   return it - localExprs.begin();
 }
 
 /// Simplify the affine expression by flattening it and reconstructing it.
 AffineExpr mlir::simplifyAffineExpr(AffineExpr expr, unsigned numDims,
                                     unsigned numSymbols) {
   // Simplify semi-affine expressions separately.
   if (!expr.isPureAffine())
     expr = simplifySemiAffine(expr);
-  if (!expr.isPureAffine())
-    return expr;
 
   SimpleAffineExprFlattener flattener(numDims, numSymbols);
   flattener.walkPostOrder(expr);
   ArrayRef<int64_t> flattenedExpr = flattener.operandExprStack.back();
-  auto simplifiedExpr =
-      getAffineExprFromFlatForm(flattenedExpr, numDims, numSymbols,
-                                flattener.localExprs, expr.getContext());
+  if (!expr.isPureAffine() &&
+      expr == getAffineExprFromFlatForm(flattenedExpr, numDims, numSymbols,
+                                        flattener.localExprs,
+                                        expr.getContext()))
+    return expr;
+  AffineExpr simplifiedExpr =
+      expr.isPureAffine()
+          ? getAffineExprFromFlatForm(flattenedExpr, numDims, numSymbols,
+                                      flattener.localExprs, expr.getContext())
+          : getSemiAffineExprFromFlatForm(flattenedExpr, numDims, numSymbols,
+                                          flattener.localExprs,
+                                          expr.getContext());
+
   flattener.operandExprStack.pop_back();
   assert(flattener.operandExprStack.empty());
 
   return simplifiedExpr;
 }

mlir/test/Dialect/Affine/simplify-affine-structures.mlir

func @test_not_trivially_true_or_false_returning_three_results() -> (index, index, index) {
   } else {
     %c3 = arith.constant 3 : index
     %c4 = arith.constant 4 : index
     %c5 = arith.constant 5 : index
     affine.yield %c3, %c4, %c5 : index, index, index
   }
   return %res#0, %res#1, %res#2 : index, index, index
 }
+
+// -----
+
+// Test simplification of semi affine expressions.
+ // CHECK-DAG: #[[MOD:.*]] = affine_map<()[s0, s1, s2, s3, s4] -> (s3 + s4 * s1 + (s0 - s1) mod s2)>
+ // CHECK-DAG: #[[FLOORDIV:.*]] = affine_map<()[s0, s1, s2, s3] -> (s0 + s3 + (s0 - s1) floordiv s2)>
+ // CHECK-DAG: #[[PRODUCT:.*]] = affine_map<()[s0, s1, s2, s3, s4] -> (s3 + s4 + (s0 - s1) * s2)>
+ // CHECK-DAG: #[[SIMPLIFIED_FLOORDIV_RHS:.*]] = affine_map<()[s0, s1, s2, s3] -> (s3 floordiv (s2 - s0 * s1))>
+ // CHECK-DAG: #[[SIMPLIFIED_CEILDIV_RHS:.*]] = affine_map<()[s0, s1, s2, s3] -> (s3 ceildiv (s2 - s0 * s1))>
+ // CHECK-DAG: #[[SIMPLIFIED_MOD_RHS:.*]] = affine_map<()[s0, s1, s2, s3] -> (s3 mod (s2 - s0 * s1))>
+ // CHECK-DAG: #[[SORTED_INDICES_EXPR:.*]] = affine_map<()[s0, s1, s2, s3, s4] -> (s2 * s0 + s2 + s3 * s0 + s3 * s1 + s3 + s4 * s1 + s4)>
+ // CHECK: func @semiaffine_simplification(%[[ARG0:.*]]: index, %[[ARG1:.*]]: index, %[[ARG2:.*]]: index, %[[ARG3:.*]]: index, %[[ARG4:.*]]: index, %[[ARG5:.*]]: index)
+ func @semiaffine_simplification(%arg0: index, %arg1: index, %arg2: index, %arg3: index, %arg4: index, %arg5: index) -> (index, index, index, index, index, index, index) {
+   %a = affine.apply affine_map<(d0, d1)[s0, s1, s2, s3] -> ((-(d1 * s0 - (s0 - s1) mod s2) + s3) + (d0 * s1 + d1 * s0))>(%arg0, %arg1)[%arg2, %arg3, %arg4, %arg5]
+   %b = affine.apply affine_map<(d0)[s0, s1, s2, s3] -> ((-(d0 * s1 - (s0 - s1) floordiv s2) + s3) + (d0 * s1 + s0))>(%arg0)[%arg1, %arg2, %arg3, %arg4]
+   %c = affine.apply affine_map<(d0)[s0, s1, s2, s3] -> ((-(s0 - (s0 - s1) * s2) + s3) + (d0 + s0))>(%arg0)[%arg1, %arg2, %arg3, %arg4]
+   %d = affine.apply affine_map<(d0)[s0, s1, s2, s3] -> (d0 floordiv (s0 - s1 * s2 + s3 - s0))>(%arg0)[%arg0, %arg1, %arg2, %arg3]
+   %e = affine.apply affine_map<(d0)[s0, s1, s2, s3] -> (d0 ceildiv (s0 - s1 * s2 + s3 - s0))>(%arg0)[%arg0, %arg1, %arg2, %arg3]
+   %f = affine.apply affine_map<(d0)[s0, s1, s2, s3] -> (d0 mod (s0 - s1 * s2 + s3 - s0))>(%arg0)[%arg0, %arg1, %arg2, %arg3]
+   %g = affine.apply affine_map<(d0, d1, d2)[s0, s1] -> (d0 + d1 * s1 + d1 + d0 * s0 + d1 * s0 + d2 * s1 + d2)>(%arg0, %arg1, %arg2)[%arg3, %arg4]
+   return %a, %b, %c, %d, %e, %f, %g : index, index, index, index, index, index, index
+ }
+ // CHECK-DAG: %[[RESULT0:.*]] = affine.apply #[[MOD]]()[%[[ARG2]], %[[ARG3]], %[[ARG4]], %[[ARG5]], %[[ARG0]]]
+ // CHECK-DAG: %[[RESULT1:.*]] = affine.apply #[[FLOORDIV]]()[%[[ARG1]], %[[ARG2]], %[[ARG3]], %[[ARG4]]]
+ // CHECK-DAG: %[[RESULT2:.*]] = affine.apply #[[PRODUCT]]()[%[[ARG1]], %[[ARG2]], %[[ARG3]], %[[ARG4]], %[[ARG0]]]
+ // CHECK-DAG: %[[RESULT3:.*]] = affine.apply #[[SIMPLIFIED_FLOORDIV_RHS]]()[%[[ARG1]], %[[ARG2]], %[[ARG3]], %[[ARG0]]]
+ // CHECK-DAG: %[[RESULT4:.*]] = affine.apply #[[SIMPLIFIED_CEILDIV_RHS]]()[%[[ARG1]], %[[ARG2]], %[[ARG3]], %[[ARG0]]]
+ // CHECK-DAG: %[[RESULT5:.*]] = affine.apply #[[SIMPLIFIED_MOD_RHS]]()[%[[ARG1]], %[[ARG2]], %[[ARG3]], %[[ARG0]]]
+ // CHECK-DAG: %[[RESULT6:.*]] = affine.apply #[[SORTED_INDICES_EXPR]]()[%[[ARG3]], %[[ARG4]], %[[ARG0]], %[[ARG1]], %[[ARG2]]]
+ // CHECK-NEXT: return %[[RESULT0]], %[[RESULT1]], %[[RESULT2]], %[[RESULT3]], %[[RESULT4]], %[[RESULT5]], %[[RESULT6]]

utils/arcanist/clang-format.sh

 # message on line 1, char 1 with file content edited using clang-format-diff.py
 #
 # We do not use git-clang-format because it wants to modify the index,
 # and arc is already holding the lock.
 #
 # We do not look for clang-format-diff or clang-format-diff.py in the PATH
 # because whether/how these are installed differs between distributions,
 # and we have an executable copy in the tree anyway.
-arc_base_commit=$(arc which --show-base)
+arc_base_commit=$(/opt/arcanist/bin/arc which --show-base)
 git diff-index -U0 "${arc_base_commit}" "${src_file}" \
   | clang/tools/clang-format/clang-format-diff.py -style file -i -p1
 
 cp -p "${src_file}" "${formatted_file}"
 cp -p "${original_file}" "${src_file}"
 if ! diff -q "${src_file}" "${formatted_file}" > /dev/null ; then
   echo "autofix"
   echo "clang-format suggested style edits found:"
   echo "1,1"  # line,char of start of replacement.
   cat "${src_file}"
   echo ">>>>"
   cat "${formatted_file}"
   echo "<<<<"
 fi