[Mlir-commits] [mlir] [mlir][Transforms][NFC] Simplify `buildUnresolvedMaterialization` implementation (PR #121651)

Matthias Springer llvmlistbot at llvm.org
Sat Jan 4 07:58:16 PST 2025


https://github.com/matthias-springer created https://github.com/llvm/llvm-project/pull/121651

The `buildUnresolvedMaterialization` implementation used to check if a materialization is necessary. A materialization is not necessary if the desired types already match the input. However, this situation can never happen: we look for mapped values with the desired type at the call sites before requesting a new unresolved materialization.

The previous implementation seemed incorrect because `buildUnresolvedMaterialization` created a mapping that is never rolled back. (When in reality that code was never executed, so it is technically not incorrect.)

Also fix a comment that in `findOrBuildReplacementValue` that was incorrect.


>From 6aabd138c6ca0ceca268e4d6bbd9195a9cf4ac65 Mon Sep 17 00:00:00 2001
From: Matthias Springer <mspringer at nvidia.com>
Date: Sat, 4 Jan 2025 16:51:39 +0100
Subject: [PATCH] [mlir][Transforms][NFC] Simplify
 `buildUnresolvedMaterialization` implementation

The `buildUnresolvedMaterialization` implementation used to check if a materialization is necessary. A materialization is not necessary if the desired types already match the input. However, this situation can never happen: we look for mapped values with the desired type at the call sites before requesting a new unresolved materialization.
---
 mlir/lib/Transforms/Utils/DialectConversion.cpp | 17 +++++------------
 1 file changed, 5 insertions(+), 12 deletions(-)

diff --git a/mlir/lib/Transforms/Utils/DialectConversion.cpp b/mlir/lib/Transforms/Utils/DialectConversion.cpp
index 0e577d2d39de3d..fbecb874f17e0a 100644
--- a/mlir/lib/Transforms/Utils/DialectConversion.cpp
+++ b/mlir/lib/Transforms/Utils/DialectConversion.cpp
@@ -1430,13 +1430,8 @@ ValueRange ConversionPatternRewriterImpl::buildUnresolvedMaterialization(
     UnrealizedConversionCastOp *castOp) {
   assert((!originalType || kind == MaterializationKind::Target) &&
          "original type is valid only for target materializations");
-
-  // Avoid materializing an unnecessary cast.
-  if (TypeRange(inputs) == outputTypes) {
-    if (!valuesToMap.empty())
-      mapping.map(std::move(valuesToMap), inputs);
-    return inputs;
-  }
+  assert(TypeRange(inputs) != outputTypes &&
+         "materialization is not necessary");
 
   // Create an unresolved materialization. We use a new OpBuilder to avoid
   // tracking the materialization like we do for other operations.
@@ -1455,7 +1450,9 @@ ValueRange ConversionPatternRewriterImpl::buildUnresolvedMaterialization(
 
 Value ConversionPatternRewriterImpl::findOrBuildReplacementValue(
     Value value, const TypeConverter *converter) {
-  // Find a replacement value with the same type.
+  // Try to find a replacement value with the same type in the conversion value
+  // mapping. This includes cached materializations. We try to reuse those
+  // instead of generating duplicate IR.
   ValueVector repl = mapping.lookupOrNull(value, value.getType());
   if (!repl.empty())
     return repl.front();
@@ -1489,10 +1486,6 @@ Value ConversionPatternRewriterImpl::findOrBuildReplacementValue(
   // in the conversion value mapping.) The insertion point of the
   // materialization must be valid for all future users that may be created
   // later in the conversion process.
-  //
-  // Note: Instead of creating new IR, `buildUnresolvedMaterialization` may
-  // return an already existing, cached materialization from the conversion
-  // value mapping.
   Value castValue =
       buildUnresolvedMaterialization(MaterializationKind::Source,
                                      computeInsertPoint(repl), value.getLoc(),



More information about the Mlir-commits mailing list