Referring Image Segmentation (RIS) requires methods to generate more expressive cross-modality feature and efficiently integrate multiscale features. Existing methods usually fuse the vision and re-aligned language features via a stack of convolutions. They do not explicitly consider modeling relationship across the two features from a structured perspective, hence weakening the global-expressive ability of the fused feature. Besides, their designing network ignores the feature propensity, severely impairing the feature reusibility and their performance. In this paper, we propose a Bidirectional Spatial Semantics Correlation (BSSC) module to overcome the mentioned problem, which replaces the traditional integration of vision and aligned language features with bidirectional slice-by-slice correlation. In BSSC, each vertical or horizontal feature slice from one modality is used to recurrently correlate the feature of other modality into a set of spatial activation maps. The module helps to activate the referent wherever it is. In addition, we propose a High-level Referent Semantics Guided decoder (HRSG). In each decoding stage, HRSG uses the referent-discernible features from previous decoding stages and sentence-level language feature to generate a pair of potent channel and spatial attentions, facilitating to purify the early vision feature before its integration at a low-cost way. Extensive results show that our proposed method performs favorably over others.