hi! On Wed, Mar 03, 2021 at 04:22:29PM -0300, Alexandre Oliva wrote: > While we could hide the bug/missing feature in add_options_for_sqrt_insn > by constraining check_effective_target_sqrt_insn, the result would be > just reduced test coverage for powerpc builds that defaulted to > -mno-powerpc-gpopt. A downside without any upside.
You just *cannot* use add_options_for_sqrt_insn usefully ever (on our configurations): unless you override which -mcpu= is used, and whether hard float is enabled, you cannot depend on having classic float (and you need to disable VSX or detect x[sv]sqrtdp etc. as well as fsqrt). > Whereas if we fix the former proc to perform its documented function on > powerpc, namely return the options required to enable the fsqrt insn, But in all configurations where you *can*, it already is, unless it is "manually" disabled by using -msoft-float. Which a) almost no one does, and b) overriding that causes *less* testing, not *more*. Anyway, this is PR99352 now, and I have a patch (that works with -msoft-float as well, etc.) Will commit later today. Segher
