[mlir] Add forall canonicalization to replace constant induction vars (#112764)

Adds a canonicalization pattern for scf.forall that replaces constant
induction variables with a constant index. There is a similar
canonicalization that completely removes constant induction variables
from the loop, but that pattern does not apply on foralls with mappings,
so this one is necessary for those cases.

---------

Signed-off-by: Max Dawkins <[email protected]>

Author: Max191

mlir/lib/Dialect/SCF/IR/SCF.cpp

@@ -1767,6 +1767,31 @@ struct ForallOpSingleOrZeroIterationDimsFolder
   }
 };
 
+/// Replace all induction vars with a single trip count with their lower bound.
+struct ForallOpReplaceConstantInductionVar : public OpRewritePattern<ForallOp> {
+  using OpRewritePattern<ForallOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(ForallOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    bool changed = false;
+    for (auto [lb, ub, step, iv] :
+         llvm::zip(op.getMixedLowerBound(), op.getMixedUpperBound(),
+                   op.getMixedStep(), op.getInductionVars())) {
+      if (iv.getUses().begin() == iv.getUses().end())
+        continue;
+      auto numIterations = constantTripCount(lb, ub, step);
+      if (!numIterations.has_value() || numIterations.value() != 1) {
+        continue;
+      }
+      rewriter.replaceAllUsesWith(
+          iv, getValueOrCreateConstantIndexOp(rewriter, loc, lb));
+      changed = true;
+    }
+    return success(changed);
+  }
+};
+
 struct FoldTensorCastOfOutputIntoForallOp
     : public OpRewritePattern<scf::ForallOp> {
   using OpRewritePattern<scf::ForallOp>::OpRewritePattern;
@@ -1851,7 +1876,8 @@ void ForallOp::getCanonicalizationPatterns(RewritePatternSet &results,
                                            MLIRContext *context) {
   results.add<DimOfForallOp, FoldTensorCastOfOutputIntoForallOp,
               ForallOpControlOperandsFolder, ForallOpIterArgsFolder,
-              ForallOpSingleOrZeroIterationDimsFolder>(context);
+              ForallOpSingleOrZeroIterationDimsFolder,
+              ForallOpReplaceConstantInductionVar>(context);
 }
 
 /// Given the region at `index`, or the parent operation if `index` is None,

mlir/test/Dialect/SCF/canonicalize.mlir

@@ -1617,7 +1617,7 @@ func.func @do_not_inline_distributed_forall_loop(
     %in: tensor<8x8xf32>) -> tensor<8x8xf32> {
   %cst = arith.constant 0.000000e+00 : f32
   %0 = tensor.empty() : tensor<8x8xf32>
-  %1 = scf.forall (%i, %j) = (0, 0) to (1, 1) step (8, 8)
+  %1 = scf.forall (%i, %j) = (0, 4) to (1, 5) step (8, 8)
       shared_outs (%out_ = %0) -> (tensor<8x8xf32>) {
     %slice = tensor.extract_slice %out_[%i, %j] [2, 3] [1, 1]
       : tensor<8x8xf32> to tensor<2x3xf32>
@@ -1632,6 +1632,8 @@ func.func @do_not_inline_distributed_forall_loop(
 }
 // CHECK-LABEL: @do_not_inline_distributed_forall_loop
 // CHECK: scf.forall
+// CHECK:   tensor.extract_slice %{{.*}}[0, 4] [2, 3] [1, 1]
+// CHECK:   tensor.parallel_insert_slice %{{.*}}[0, 4] [2, 3] [1, 1]
 
 // -----