The attached source file contains an implementation of convolution of integer values. It has one loop that makes 4 calls to an inlined function. Build it with version 7.4.24 ...

% cl6x -@options.txt file.c

Inspect the resulting assembly code, and see the loop gets II=4. Build it again with 8.3.12, and the loop gets II=7.

Somehow, version 7.4.24 optimizes out some short-lived variables in the inlined function. They do not appear in the optimizer generated comments. But 8.3.12 has them. That generates more MV and ADD instructions in the loop, driving up the resource bound.