[mlir][arith] arith-to-apfloat: Bail on unsupported bitwidth

[matthias-springer]

Bitwidths greater than 64 are not supported by arith-to-apfloat.

[llvmbot]

@llvm/pr-subscribers-mlir

Author: Matthias Springer (matthias-springer)

Changes

Bitwidths greater than 64 are not supported by arith-to-apfloat.

---

Full diff: https://github.com/llvm/llvm-project/pull/170994.diff

2 Files Affected:

• (modified) mlir/lib/Conversion/ArithToAPFloat/ArithToAPFloat.cpp (+26-8)
• (modified) mlir/test/Conversion/ArithToApfloat/arith-to-apfloat.mlir (+25)

diff --git a/mlir/lib/Conversion/ArithToAPFloat/ArithToAPFloat.cpp b/mlir/lib/Conversion/ArithToAPFloat/ArithToAPFloat.cpp
index 5c68236526b7d..4776ba0f49b94 100644
--- a/mlir/lib/Conversion/ArithToAPFloat/ArithToAPFloat.cpp
+++ b/mlir/lib/Conversion/ArithToAPFloat/ArithToAPFloat.cpp
@@ -102,6 +102,10 @@ struct BinaryArithOpToAPFloatConversion final : OpRewritePattern<OpTy> {
 
   LogicalResult matchAndRewrite(OpTy op,
                                 PatternRewriter &rewriter) const override {
+    if (op.getType().getIntOrFloatBitWidth() > 64)
+      return rewriter.notifyMatchFailure(op,
+                                         "bitwidth > 64 bits is not supported");
+
     // Get APFloat function from runtime library.
     FailureOr<FuncOp> fn =
         lookupOrCreateBinaryFn(rewriter, symTable, APFloatName);
@@ -148,6 +152,11 @@ struct FpToFpConversion final : OpRewritePattern<OpTy> {
 
   LogicalResult matchAndRewrite(OpTy op,
                                 PatternRewriter &rewriter) const override {
+    if (op.getType().getIntOrFloatBitWidth() > 64 ||
+        op.getOperand().getType().getIntOrFloatBitWidth() > 64)
+      return rewriter.notifyMatchFailure(op,
+                                         "bitwidth > 64 bits is not supported");
+
     // Get APFloat function from runtime library.
     auto i32Type = IntegerType::get(symTable->getContext(), 32);
     auto i64Type = IntegerType::get(symTable->getContext(), 64);
@@ -195,9 +204,10 @@ struct FpToIntConversion final : OpRewritePattern<OpTy> {
 
   LogicalResult matchAndRewrite(OpTy op,
                                 PatternRewriter &rewriter) const override {
-    if (op.getType().getIntOrFloatBitWidth() > 64)
-      return rewriter.notifyMatchFailure(
-          op, "result type > 64 bits is not supported");
+    if (op.getType().getIntOrFloatBitWidth() > 64 ||
+        op.getOperand().getType().getIntOrFloatBitWidth() > 64)
+      return rewriter.notifyMatchFailure(op,
+                                         "bitwidth > 64 bits is not supported");
 
     // Get APFloat function from runtime library.
     auto i1Type = IntegerType::get(symTable->getContext(), 1);
@@ -252,11 +262,10 @@ struct IntToFpConversion final : OpRewritePattern<OpTy> {
 
   LogicalResult matchAndRewrite(OpTy op,
                                 PatternRewriter &rewriter) const override {
-    Location loc = op.getLoc();
-    if (op.getIn().getType().getIntOrFloatBitWidth() > 64) {
-      return rewriter.notifyMatchFailure(
-          loc, "integer bitwidth > 64 is not supported");
-    }
+    if (op.getType().getIntOrFloatBitWidth() > 64 ||
+        op.getOperand().getType().getIntOrFloatBitWidth() > 64)
+      return rewriter.notifyMatchFailure(op,
+                                         "bitwidth > 64 bits is not supported");
 
     // Get APFloat function from runtime library.
     auto i1Type = IntegerType::get(symTable->getContext(), 1);
@@ -270,6 +279,7 @@ struct IntToFpConversion final : OpRewritePattern<OpTy> {
 
     rewriter.setInsertionPoint(op);
     // Cast operands to 64-bit integers.
+    Location loc = op.getLoc();
     auto inIntTy = cast<IntegerType>(op.getOperand().getType());
     Value operandBits = op.getOperand();
     if (operandBits.getType().getIntOrFloatBitWidth() < 64) {
@@ -317,6 +327,10 @@ struct CmpFOpToAPFloatConversion final : OpRewritePattern<arith::CmpFOp> {
 
   LogicalResult matchAndRewrite(arith::CmpFOp op,
                                 PatternRewriter &rewriter) const override {
+    if (op.getLhs().getType().getIntOrFloatBitWidth() > 64)
+      return rewriter.notifyMatchFailure(op,
+                                         "bitwidth > 64 bits is not supported");
+
     // Get APFloat function from runtime library.
     auto i1Type = IntegerType::get(symTable->getContext(), 1);
     auto i8Type = IntegerType::get(symTable->getContext(), 8);
@@ -456,6 +470,10 @@ struct NegFOpToAPFloatConversion final : OpRewritePattern<arith::NegFOp> {
 
   LogicalResult matchAndRewrite(arith::NegFOp op,
                                 PatternRewriter &rewriter) const override {
+    if (op.getOperand().getType().getIntOrFloatBitWidth() > 64)
+      return rewriter.notifyMatchFailure(op,
+                                         "bitwidth > 64 bits is not supported");
+
     // Get APFloat function from runtime library.
     auto i32Type = IntegerType::get(symTable->getContext(), 32);
     auto i64Type = IntegerType::get(symTable->getContext(), 64);
diff --git a/mlir/test/Conversion/ArithToApfloat/arith-to-apfloat.mlir b/mlir/test/Conversion/ArithToApfloat/arith-to-apfloat.mlir
index 950d2cecefa95..ab05edebec71d 100644
--- a/mlir/test/Conversion/ArithToApfloat/arith-to-apfloat.mlir
+++ b/mlir/test/Conversion/ArithToApfloat/arith-to-apfloat.mlir
@@ -263,3 +263,28 @@ func.func @maxnumf(%arg0: f32, %arg1: f32) {
   %0 = arith.maxnumf %arg0, %arg1 : f32
   return
 }
+
+// -----
+
+// CHECK-LABEL: func.func @unsupported_bitwidth
+//       CHECK:   arith.addf {{.*}} : f128
+//       CHECK:   arith.negf {{.*}} : f128
+//       CHECK:   arith.cmpf {{.*}} : f128
+//       CHECK:   arith.extf {{.*}} : f32 to f128
+//       CHECK:   arith.truncf {{.*}} : f128 to f32
+//       CHECK:   arith.fptosi {{.*}} : f128 to i32
+//       CHECK:   arith.fptosi {{.*}} : f32 to i92
+//       CHECK:   arith.sitofp {{.*}} : i1 to f128
+//       CHECK:   arith.sitofp {{.*}} : i92 to f32
+func.func @unsupported_bitwidth(%arg0: f128, %arg1: f128, %arg2: f32) {
+  %0 = arith.addf %arg0, %arg1 : f128
+  %1 = arith.negf %arg0 : f128
+  %2 = arith.cmpf "ult", %arg0, %arg1 : f128
+  %3 = arith.extf %arg2 : f32 to f128
+  %4 = arith.truncf %arg0 : f128 to f32
+  %5 = arith.fptosi %arg0 : f128 to i32
+  %6 = arith.fptosi %arg2 : f32 to i92
+  %7 = arith.sitofp %2 : i1 to f128
+  %8 = arith.sitofp %6 : i92 to f32
+  return
+}