[Mlir-commits] [mlir] [mlir][arith] wide integer emulation support for fpto*i ops (PR #132375)

[Jakub Kuderski]

================
@@ -974,6 +974,126 @@ struct ConvertUIToFP final : OpConversionPattern<arith::UIToFPOp> {
   }
 };
 
+//===----------------------------------------------------------------------===//
+// ConvertFPToSI
+//===----------------------------------------------------------------------===//
+
+struct ConvertFPToSI final : OpConversionPattern<arith::FPToSIOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(arith::FPToSIOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    /* Get the input float type */
+    auto inFp = adaptor.getIn();
+    auto fpTy = inFp.getType();
+    auto fpElemTy = getElementTypeOrSelf(fpTy);
+
+    Type intTy = op.getType();
+    unsigned oldBitWidth = getElementTypeOrSelf(intTy).getIntOrFloatBitWidth();
+
+    auto newTy = getTypeConverter()->convertType<VectorType>(intTy);
+    if (!newTy)
+      return rewriter.notifyMatchFailure(
+          loc, llvm::formatv("unsupported type: {0}", intTy));
+
+    /*
+    Work on the absolute value and then convert the result to signed integer.
+    Defer absolute value to fptoui. If minSInt < fp < maxSInt, i.e.
+    if the fp is representable in signed i2N, emits the correct result.
+    Else, the result is UB.
+    */
----------------
kuhar wrote:

Do not use C block comments.

https://github.com/llvm/llvm-project/pull/132375