Andrew Pinski <pins...@gmail.com> wrote: > I am trying to understand the difference between this option and > -mgeneral-regs-only since you mentioned that FPU will be disabled for > those cases.
Thanks for taking a look at the patch; I was also surprised that there wasn't some existing option that I could use. This new option only disables FPU registers when used implicitly by memory options, such as struct copies and calls to memset, memmove or memcpy. Unlike -mgeneral-regs-only, other uses, like explicit floating point or SIMD operations will still use those registers. The goal is to let code which doesn't "obviously" use the FPU avoid unintentional usage. These memory operations were the only place I found which used them for non-SIMD/FP data. > Also I am not sure aarch64-elf should turn off this by default since > there are some bare metal env which support the full SIMD registers > for interrupt save/restoring. The impetus for this series was use of va_start from an interrupt handler in Zephyr. Zephyr fully supports save/restore of SIMD registers in task context, but if used in an interrupt handler, the OS needs to disable interrupts as there's no place to save the registers in a nested handler. So, the existing compiler generates code that "mostly works", but it changes the semantics of interrupt processing. Fortunately, the Zephyr test suite includes an explicit check to verify correct operation of nested interrupt handlers which failed on aarch64 with gcc 14.3. The previous compiler version in use by Zephyr (12.2) didn't use SIMD registers for these memory operations, so this test appears as a regression. -- -keith
signature.asc
Description: PGP signature
