[Mlir-commits] [mlir] [MLIR][Arith] Add canonicalization rules for int-to-float of integer extension (PR #185386)

Adam Paszke llvmlistbot at llvm.org
Tue Mar 10 04:13:17 PDT 2026


https://github.com/apaszke updated https://github.com/llvm/llvm-project/pull/185386

>From 7e22ea6764320cff996c75fa6ef9dee5501ac440 Mon Sep 17 00:00:00 2001
From: Adam Paszke <apaszke at google.com>
Date: Mon, 9 Mar 2026 10:13:50 +0000
Subject: [PATCH] [MLIR][Arith] Add canonicalization rules for int-to-float of
 integer extension
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Three patterns are valid but were missing:

1. `sitofp(extsi(x)) → sitofp(x)`:
   extsi preserves the sign and value, so it represents the
   same signed integer as x.

2. `uitofp(extui(x)) → uitofp(x)`:
   same reasoning as above, but for unsigned extension.

3. sitofp(extui(x)) → uitofp(x)
   extui zero-extends, so the extended value is always non-negative.
   For non-negative integers, sitofp and uitofp produce the same result,
   meaning we could replace the left expression by `uitofp(extui(x))`. At
   this point rule 2. above can be used to simplify further to `uitofp(x)`.
---
 .../include/mlir/Dialect/Arith/IR/ArithOps.td |  2 +
 .../Dialect/Arith/IR/ArithCanonicalization.td | 23 ++++++++++++
 mlir/lib/Dialect/Arith/IR/ArithOps.cpp        | 10 +++++
 mlir/test/Dialect/Arith/canonicalize.mlir     | 37 ++++++++++++++++++-
 .../emulate-wide-int-canonicalization.mlir    | 11 ++++--
 5 files changed, 78 insertions(+), 5 deletions(-)

diff --git a/mlir/include/mlir/Dialect/Arith/IR/ArithOps.td b/mlir/include/mlir/Dialect/Arith/IR/ArithOps.td
index 27d3d0d9b6e71..4b830c05bf585 100644
--- a/mlir/include/mlir/Dialect/Arith/IR/ArithOps.td
+++ b/mlir/include/mlir/Dialect/Arith/IR/ArithOps.td
@@ -1530,6 +1530,7 @@ def Arith_UIToFPOp :
     $in (`nneg` $nonNeg^)? attr-dict `:` type($in) `to` type($out)
   }];
   let hasFolder = 1;
+  let hasCanonicalizer = 1;
 }
 
 //===----------------------------------------------------------------------===//
@@ -1545,6 +1546,7 @@ def Arith_SIToFPOp : Arith_IToFCastOp<"sitofp"> {
     elementwise.
   }];
   let hasFolder = 1;
+  let hasCanonicalizer = 1;
 }
 
 //===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/Arith/IR/ArithCanonicalization.td b/mlir/lib/Dialect/Arith/IR/ArithCanonicalization.td
index fb9c16db91431..e22fc1d478e4f 100644
--- a/mlir/lib/Dialect/Arith/IR/ArithCanonicalization.td
+++ b/mlir/lib/Dialect/Arith/IR/ArithCanonicalization.td
@@ -421,6 +421,29 @@ def TruncFUIToFPToUIToFP :
         (Arith_UIToFPOp $x, $nneg),
         [(Constraint<CPred<"$0 == nullptr">, "default rounding mode"> $rmf)]>;
 
+//===----------------------------------------------------------------------===//
+// SIToFPOp
+//===----------------------------------------------------------------------===//
+
+// sitofp(extsi(x)) -> sitofp(x)
+def SIToFPOfExtSI :
+    Pat<(Arith_SIToFPOp (Arith_ExtSIOp $x)),
+        (Arith_SIToFPOp $x)>;
+
+// sitofp(extui(x)) -> uitofp(x)
+def SIToFPOfExtUI :
+    Pat<(Arith_SIToFPOp (Arith_ExtUIOp $x, $nneg)),
+        (Arith_UIToFPOp $x, $nneg)>;
+
+//===----------------------------------------------------------------------===//
+// UIToFPOp
+//===----------------------------------------------------------------------===//
+
+// uitofp(extui(x)) -> uitofp(x)
+def UIToFPOfExtUI :
+    Pat<(Arith_UIToFPOp (Arith_ExtUIOp $x, $nneg1), $nneg2),
+        (Arith_UIToFPOp $x, $nneg1)>;
+
 //===----------------------------------------------------------------------===//
 // MulFOp
 //===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/Arith/IR/ArithOps.cpp b/mlir/lib/Dialect/Arith/IR/ArithOps.cpp
index 73fa0df80bcba..6f368604df65a 100644
--- a/mlir/lib/Dialect/Arith/IR/ArithOps.cpp
+++ b/mlir/lib/Dialect/Arith/IR/ArithOps.cpp
@@ -1756,6 +1756,11 @@ OpFoldResult arith::UIToFPOp::fold(FoldAdaptor adaptor) {
       });
 }
 
+void arith::UIToFPOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
+                                                  MLIRContext *context) {
+  patterns.add<UIToFPOfExtUI>(context);
+}
+
 //===----------------------------------------------------------------------===//
 // SIToFPOp
 //===----------------------------------------------------------------------===//
@@ -1778,6 +1783,11 @@ OpFoldResult arith::SIToFPOp::fold(FoldAdaptor adaptor) {
       });
 }
 
+void arith::SIToFPOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
+                                                  MLIRContext *context) {
+  patterns.add<SIToFPOfExtSI, SIToFPOfExtUI>(context);
+}
+
 //===----------------------------------------------------------------------===//
 // FPToUIOp
 //===----------------------------------------------------------------------===//
diff --git a/mlir/test/Dialect/Arith/canonicalize.mlir b/mlir/test/Dialect/Arith/canonicalize.mlir
index 035c10e78bf9b..26b04e4209a43 100644
--- a/mlir/test/Dialect/Arith/canonicalize.mlir
+++ b/mlir/test/Dialect/Arith/canonicalize.mlir
@@ -875,7 +875,42 @@ func.func @truncUitofp_nneg(%arg0: i32) -> f32 {
   return %trunc : f32
 }
 
-// TODO: We should also add a test for not folding arith.extf on information loss.
+// CHECK-LABEL: @sitofpExtsi
+//       CHECK:     %[[SITOFP:.*]] = arith.sitofp %[[ARG0:.*]] : i8 to bf16
+//       CHECK:     return %[[SITOFP]]
+func.func @sitofpExtsi(%arg0: i8) -> bf16 {
+  %extsi = arith.extsi %arg0 : i8 to i32
+  %sitofp = arith.sitofp %extsi : i32 to bf16
+  return %sitofp : bf16
+}
+
+// CHECK-LABEL: @sitofpExtui
+//       CHECK:     %[[UITOFP:.*]] = arith.uitofp %[[ARG0:.*]] : i4 to bf16
+//       CHECK-NOT: sitofp
+//       CHECK:     return %[[UITOFP]]
+func.func @sitofpExtui(%arg0: i4) -> bf16 {
+  %extui = arith.extui %arg0 : i4 to i8
+  %sitofp = arith.sitofp %extui : i8 to bf16
+  return %sitofp : bf16
+}
+
+// CHECK-LABEL: @uitofpExtui
+//       CHECK:     %[[UITOFP:.*]] = arith.uitofp %[[ARG0:.*]] : i8 to bf16
+//       CHECK:     return %[[UITOFP]]
+func.func @uitofpExtui(%arg0: i8) -> bf16 {
+  %extui = arith.extui %arg0 : i8 to i32
+  %uitofp = arith.uitofp %extui : i32 to bf16
+  return %uitofp : bf16
+}
+
+// CHECK-LABEL: @sitofpExtui_nneg
+//       CHECK:     %[[UITOFP:.*]] = arith.uitofp %[[ARG0:.*]] nneg : i4 to bf16
+//       CHECK:     return %[[UITOFP]]
+func.func @sitofpExtui_nneg(%arg0: i4) -> bf16 {
+  %extui = arith.extui %arg0 nneg : i4 to i8
+  %sitofp = arith.sitofp %extui : i8 to bf16
+  return %sitofp : bf16
+}
 // This may happen when extending f8E5M2FNUZ to f16.
 
 // CHECK-LABEL: @truncConstant
diff --git a/mlir/test/Dialect/Arith/emulate-wide-int-canonicalization.mlir b/mlir/test/Dialect/Arith/emulate-wide-int-canonicalization.mlir
index 0c95ab8284afa..75fd9eda08782 100644
--- a/mlir/test/Dialect/Arith/emulate-wide-int-canonicalization.mlir
+++ b/mlir/test/Dialect/Arith/emulate-wide-int-canonicalization.mlir
@@ -1,14 +1,17 @@
 // RUN: mlir-opt --arith-emulate-wide-int="widest-int-supported=32" --canonicalize %s | FileCheck %s
 
-// Check that we can fold away the 'hi' part calculation when it is know to be zero.
+// Check that we can fold away the 'hi' part calculation when it is known to be zero.
 //
 // CHECK-LABEL: func @uitofp_i16_ext_f64
 // CHECK-SAME:    ([[ARG:%.+]]: i16) -> f64
-// CHECK-NEXT:    [[EXT:%.+]] = arith.extui [[ARG]] : i16 to i32
-// CHECK-NEXT:    [[FP:%.+]]  = arith.uitofp [[EXT]] : i32 to f64
+// CHECK:         [[EXT:%.+]] = arith.extui [[ARG]] : i16 to i32
+// CHECK:         [[ORI:%.+]] = arith.ori [[EXT]], {{.*}} : i32
+// CHECK:         [[FP:%.+]]  = arith.uitofp [[ORI]] : i32 to f64
 // CHECK-NEXT:    return [[FP]] : f64
 func.func @uitofp_i16_ext_f64(%a : i16) -> f64 {
   %ext = arith.extui %a : i16 to i64
-  %r = arith.uitofp %ext : i64 to f64
+  %c = arith.constant 1 : i64
+  %or = arith.ori %ext, %c : i64
+  %r = arith.uitofp %or : i64 to f64
   return %r : f64
 }



More information about the Mlir-commits mailing list