[mlir][transform] Update transform.loop.peel

[banach-space]

This patch updates transform.loop.peel so that this Op returns two
rather than one handle:

• one for the peeled loop, and
• one for the remainder loop.

Also, following this change this Op will fail if peeling fails. This is
consistent with other similar Ops that also fail if no transformation
takes place.

[llvmbot]

@llvm/pr-subscribers-mlir-scf
@llvm/pr-subscribers-mlir

@llvm/pr-subscribers-mlir-linalg

Changes

This patch updates transform.loop.peel so that this Op returns two
rather than one handle:

• one for the peeled loop, and
• one for the remainder loop.

Also, following this change this Op will fail if peeling fails. This is
consistent with other similar Ops that also fail if no transformation
takes place.

---

Full diff: https://github.com/llvm/llvm-project/pull/67482.diff

5 Files Affected:

• (modified) mlir/include/mlir/Dialect/SCF/TransformOps/SCFTransformOps.td (+9-10)
• (modified) mlir/lib/Dialect/SCF/TransformOps/SCFTransformOps.cpp (+8-6)
• (modified) mlir/test/Dialect/Linalg/transform-op-fuse.mlir (+1-1)
• (modified) mlir/test/Dialect/SCF/transform-ops-invalid.mlir (+20)
• (modified) mlir/test/Dialect/SCF/transform-ops.mlir (+1-1)

diff --git a/mlir/include/mlir/Dialect/SCF/TransformOps/SCFTransformOps.td b/mlir/include/mlir/Dialect/SCF/TransformOps/SCFTransformOps.td
index 207a004c54ef5af..3adf5f0dc97dd5a 100644
--- a/mlir/include/mlir/Dialect/SCF/TransformOps/SCFTransformOps.td
+++ b/mlir/include/mlir/Dialect/SCF/TransformOps/SCFTransformOps.td
@@ -142,23 +142,22 @@ def LoopPeelOp : Op<Transform_Dialect, "loop.peel",
 
      This operation ignores non-scf::ForOp ops and drops them in the return.
 
-     This operation always succeeds and returns the scf::ForOp with the
-     postcondition: "the loop trip count is divisible by the step".
-     This operation may return the same unmodified loop handle when peeling did
-     not modify the IR (i.e. the loop trip count was already divisible).
+     This operation returns two scf::ForOp, with the first Op satisfying the
+     postcondition: "the loop trip count is divisible by the step". The second
+     loop Op contains the remaining iteration. Note that even though the
+     Payload IR modification may be performed in-place, this operation consumes
+     the operand handle and produces a new one.
 
-     Note that even though the Payload IR modification may be performed
-     in-place, this operation consumes the operand handle and produces a new
-     one.
+     #### Return Modes
 
-     TODO: Return both the peeled loop and the remainder loop.
+     Produces a definite failure if peeling fails.
   }];
 
   let arguments =
       (ins Transform_ScfForOp:$target,
            DefaultValuedAttr<BoolAttr, "false">:$fail_if_already_divisible);
-  // TODO: Return both the peeled loop and the remainder loop.
-  let results = (outs TransformHandleTypeInterface:$transformed);
+  let results = (outs TransformHandleTypeInterface:$peeled_loop,
+                      TransformHandleTypeInterface:$remainder_loop);
 
   let assemblyFormat =
     "$target attr-dict `:` functional-type(operands, results)";
diff --git a/mlir/lib/Dialect/SCF/TransformOps/SCFTransformOps.cpp b/mlir/lib/Dialect/SCF/TransformOps/SCFTransformOps.cpp
index d7e8c38478ced1a..65d503d7c4ad8b8 100644
--- a/mlir/lib/Dialect/SCF/TransformOps/SCFTransformOps.cpp
+++ b/mlir/lib/Dialect/SCF/TransformOps/SCFTransformOps.cpp
@@ -226,14 +226,16 @@ transform::LoopPeelOp::applyToOne(transform::TransformRewriter &rewriter,
                                   transform::ApplyToEachResultList &results,
                                   transform::TransformState &state) {
   scf::ForOp result;
-  // This helper returns failure when peeling does not occur (i.e. when the IR
-  // is not modified). This is not a failure for the op as the postcondition:
-  //    "the loop trip count is divisible by the step"
-  // is valid.
   LogicalResult status =
       scf::peelForLoopAndSimplifyBounds(rewriter, target, result);
-  // TODO: Return both the peeled loop and the remainder loop.
-  results.push_back(failed(status) ? target : result);
+  if (failed(status)) {
+    DiagnosedSilenceableFailure diag = emitSilenceableError()
+                                       << "failed to peel";
+    return diag;
+  }
+  results.push_back(target);
+  results.push_back(result);
+
   return DiagnosedSilenceableFailure::success();
 }
 
diff --git a/mlir/test/Dialect/Linalg/transform-op-fuse.mlir b/mlir/test/Dialect/Linalg/transform-op-fuse.mlir
index d3772e9c04eef52..72c61661b55efd3 100644
--- a/mlir/test/Dialect/Linalg/transform-op-fuse.mlir
+++ b/mlir/test/Dialect/Linalg/transform-op-fuse.mlir
@@ -48,7 +48,7 @@ transform.sequence failures(propagate) {
   %0 = transform.structured.match ops{["linalg.elemwise_binary"]} in %arg1 : (!transform.any_op) -> !transform.any_op
   %1, %loops:2 = transform.structured.fuse %0 {tile_sizes = [32, 32], tile_interchange = [0, 1]}
     : (!transform.any_op) -> (!transform.any_op, !transform.op<"scf.for">, !transform.any_op)
-  transform.loop.peel %loops#0 : (!transform.op<"scf.for">) -> !transform.any_op
+  transform.loop.peel %loops#0 : (!transform.op<"scf.for">) -> (!transform.any_op, !transform.any_op)
 }
 
 // -----
diff --git a/mlir/test/Dialect/SCF/transform-ops-invalid.mlir b/mlir/test/Dialect/SCF/transform-ops-invalid.mlir
index b69ea0f5975cd80..e8212f500e693ad 100644
--- a/mlir/test/Dialect/SCF/transform-ops-invalid.mlir
+++ b/mlir/test/Dialect/SCF/transform-ops-invalid.mlir
@@ -59,3 +59,23 @@ transform.sequence failures(propagate) {
   // expected-error @below {{failed to outline}}
   transform.loop.outline %0 {func_name = "foo"} : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
 }
+
+// -----
+
+func.func @test_loops_do_not_get_peeled() {
+  %lb = arith.constant 0 : index
+  %ub = arith.constant 40 : index
+  %step = arith.constant 5 : index
+  scf.for %i = %lb to %ub step %step {
+    arith.addi %i, %i : index
+  }
+  return
+}
+
+transform.sequence failures(propagate) {
+^bb1(%arg1: !transform.any_op):
+  %0 = transform.structured.match ops{["arith.addi"]} in %arg1 : (!transform.any_op) -> !transform.any_op
+  %1 = transform.loop.get_parent_for %0 : (!transform.any_op) -> !transform.op<"scf.for">
+  // expected-error @below {{failed to peel}}
+  transform.loop.peel %1 : (!transform.op<"scf.for">) -> (!transform.any_op, !transform.any_op)
+}
diff --git a/mlir/test/Dialect/SCF/transform-ops.mlir b/mlir/test/Dialect/SCF/transform-ops.mlir
index 7718ebcd1fc6ae0..779e750fb500326 100644
--- a/mlir/test/Dialect/SCF/transform-ops.mlir
+++ b/mlir/test/Dialect/SCF/transform-ops.mlir
@@ -107,7 +107,7 @@ transform.sequence failures(propagate) {
 ^bb1(%arg1: !transform.any_op):
   %0 = transform.structured.match ops{["arith.addi"]} in %arg1 : (!transform.any_op) -> !transform.any_op
   %1 = transform.loop.get_parent_for %0 : (!transform.any_op) -> !transform.op<"scf.for">
-  transform.loop.peel %1 : (!transform.op<"scf.for">) -> !transform.any_op
+  transform.loop.peel %1 : (!transform.op<"scf.for">) -> (!transform.any_op, !transform.any_op)
 }
 
 // -----

[ftynse]

Could we double-check that if one iteration is peeled off, the loop is still present in the IR, and the handle expectedly points to that loop?

Also FYI, I quickly looked at the internals of the peeling implementation and it mixes Rewriter-based and direct IR mutation, which may lead to crashes that are fun to debug.

[banach-space]

Could we double-check that if one iteration is peeled off, the loop is still present in the IR, and the handle expectedly points to that loop?

Sure, added extra verification in a follow-up fix-up commit.

Also FYI, I quickly looked at the internals of the peeling implementation and it mixes Rewriter-based and direct IR mutation, which may lead to crashes that are fun to debug.

Thanks for the heads-up! ATM I am using this to play around with vectorisation of the remainder loop and it works fine ™️ 🤞🏻 .

[aartbik]

This broke a bot:

Failed Tests (1):
MLIR :: python/dialects/transform_loop_ext.py

[banach-space]

This broke a bot:

Failed Tests (1): MLIR :: python/dialects/transform_loop_ext.py

Sorry about that, patch reverted. I forgot about the Python bindings :(