Macro-operation fusion represents an advanced approach to optimizing instruction streams with the goal of enhancing processor efficiency, offering potential benefits for compute-intensive domains including deep learning and large language models. This work presents a software-based framework designed to systematically explore instruction fusion strategies targeting the RISC-V ISA for TensorFlow-based models. The framework combines model transformation, dynamic instruction trace analysis, and fusion simulation to identify and evaluate both established and novel fusion idioms targeting integer and floating-point operations. Evaluation across a diverse set of benchmarks, spanning classical computational tasks and modern deep learning workloads, demonstrates that the proposed fusion techniques effectively fuse, on average, 30% of dynamic instructions in deep learning models, and over 60% fusion in large language models such as GPT-2. These results provide valuable insights into the trade-offs between software-driven optimization and necessary hardware extensions, informing future RISC-V microarchitecture and compiler design aimed at maximizing instruction-level parallelism.