Index: flang/lib/Optimizer/CodeGen/CodeGen.cpp
===================================================================
--- flang/lib/Optimizer/CodeGen/CodeGen.cpp
+++ flang/lib/Optimizer/CodeGen/CodeGen.cpp
@@ -3772,6 +3772,11 @@
     mathConvertionPM.addPass(
         mlir::createConvertMathToFuncs(mathToFuncsOptions));
     mathConvertionPM.addPass(mlir::createConvertComplexToStandardPass());
+    // Convert Math dialect operations into LLVM dialect operations.
+    // There is no way to prefer MathToLLVM patterns over MathToLibm
+    // patterns (applied below), so we have to run MathToLLVM conversion here.
+    mathConvertionPM.addNestedPass<mlir::func::FuncOp>(
+        mlir::createConvertMathToLLVMPass());
     if (mlir::failed(runPipeline(mathConvertionPM, mod)))
       return signalPassFailure();
 
@@ -3826,9 +3831,8 @@
     mlir::arith::populateArithToLLVMConversionPatterns(typeConverter, pattern);
     mlir::cf::populateControlFlowToLLVMConversionPatterns(typeConverter,
                                                           pattern);
-    // Convert math-like dialect operations, which can be produced
-    // when late math lowering mode is used, into llvm dialect.
-    mlir::populateMathToLLVMConversionPatterns(typeConverter, pattern);
+    // Math operations that have not been converted yet must be converted
+    // to Libm.
     mlir::populateMathToLibmConversionPatterns(pattern);
     mlir::populateComplexToLLVMConversionPatterns(typeConverter, pattern);
     mlir::ConversionTarget target{*context};