dtrace tests could stop fighting the compiler
Rather than entering an arms race with the optimizers of various compilers, which due to the small and contained nature of the C code in the dtrace tests tend to have a field day, it is perhaps a better idea to just cease optimizing them.
Most of the tests care solely about ABI-related things (arguments, call/return, etc) which are actually more likely to break in an optimized potted test (and, in fact, do). The tests which care about the actual instruction stream are limited to tests with implement every insn in a given small program, one to test doing so with gcc-compiled executables, the other to test the same when using the FPU. Neither of these is affected in purpose by the optimizer.
That said, someone with more expertise should have the opportunity to say that this is foolish, and we should do whatever is necessary to win out over the optimizer while retaining the optimization level.