Evaluating the effectiveness of loop transformation optimizations requires benchmark suites that cover diverse loop properties and offer rich transformation opportunities. However, existing loop code datasets and generators face significant limitations in producing code with varied loop properties, particularly in generating realistic dependence patterns and complex loop structures. These limitations diminish their utility for comprehensive compiler testing and for training Large Language Models in code optimization. In this paper, we present PLCG, an enhanced parametric loop code generator that systematically produces diverse yet semantically legal loop code for optimization benchmarking. Building on parameter-driven approaches, PLCG introduces three key innovations: (1) a comprehensive parameter-to-property generation framework that translates high-level parameters into loop properties with array cross-statement coordination, diversified access functions, and flexible iteration domain; (2) a hybrid dependence generation strategy that combines parent-to-child and child-to-parent strategies to control dependence patterns while preventing illegal cycles; and (3) an adaptive parameter resolution module that employs weighted mechanisms for selecting dependence candidates, establishing iterator mappings, and assigning dependence distance values. We evaluate PLCG against state-of-the-art C loop code generators, including COLA-Gen, YARPGen, and LOOPRAG, across multiple dimensions: loop property diversity, loop transformation-triggering capability, and similarity to real-world code. Among twelve key loop properties, PLCG achieves five best-in-class metrics and six second-best metrics. In terms of transformation coverage, PLCG is the only generator that triggers all seven classical loop transformations, showing clear advantages in distribution and reversal, while delivering comparable results in the remaining transformations.

Multitask learning (MTL) is a joint learning paradigm, which fuses multiple related tasks together to achieve the better performance than single-task learning methods. It has been observed by many researchers that different tasks with certain similarities share a low-dimensional common yet latent subspace. In order to get the low-rank structure shared across tasks, trace norm has been used as a convex relaxation of the rank minimization problem. However, trace norm is not a tight approximation for the rank function. To address this important issue, we propose two novel regularization-based models to approximate the rank minimization problem by minimizing the k minimal singular values. For our new models, if the minimal singular values are suppressed to zeros, the rank would also be reduced. Compared with the standard trace norm, our new regularization-based models are the tighter approximations, which can help our models capture the low-dimensional subspace among multiple tasks better. Besides, it is an NP-hard problem to directly solve the exact rank minimization problem for our models. In this article, we proposed two simple but effective strategies to optimize our models, which tactically solves the exact rank minimization problem by setting a large penalizing parameter. Experimental results performed on synthetic and real-world benchmark datasets demonstrate that the proposed models have the ability of learning the low-rank structure shared across tasks and the better performance than other classical MTL methods.