Compiler fuzzing is one of the commonly used techniques to test the functionality and safety of compilers. The fuzzer produces grammatically correct test cases to test the deep parts of the compiler. Recently, recurrent neural networks-based deep learning methods have been introduced to the test case generation process. Aiming at the problems of insufficient grammatical accuracy and low generation efficiency when generating test cases, a method for generating compiler fuzzing test cases is proposed based on feed-forward neural networks, and the prototype tool FAIR is designed and implemented. Different from the method based on token sequence learning, FAIR extracts code fragments from the abstract syntax tree, and uses a self-attention-based feed-forward neural network to capture the grammatical associations between code fragments. After learning a generative model of the programming language, fair automatically produce diverse test cases. Experimental results show that FAIR is superior to its competitors in terms of grammatical accuracy and generation efficiency of generating test cases. The proposed method has significantly improved the ability to detect compiler software defects, and has successfully detected 20 software defects in GCC and LLVM. In addition, the method has soundportability. The simple ported FAIR-JS has detected 2 defects in the JavaScript engine.