When generating floating-point code, clang and LLVM will currently assume that the program always operates under default floating-point control modes, i.e. using the default rounding mode and with floating-point exceptions disabled, and never checks the floating-point status flags. This means that code that does attempt to make use of these IEEE features will not work reliably. The C standard defines a pragma FENV_ACCESS that is intended to instruct the compiler to switch to a method of generating code that will allow these features to be used, but this pragma and the associated infrastructure is not yet implemented in clang and LLVM.
The purpose of this BoF is to bring together all parties interested in this feature, whether as potential users, or as experts in any of the parts of the compiler that will need to be modified to implement it, from the clang front end, through the optimizers, to the various back ends that need to emit appropriate code for their platform. We will discuss the current status of the partial infrastructure that is already present, identify the pieces that are still missing, and hopefully agree on next steps to move towards a full implementation of pragma FENV_ACCESS in clang and LLVM.
