On Tue, 10 May 2022, Christophe Lyon via Gcc-patches wrote: > > Note that for conversion from DFP to HFmode, double rounding (going via > > SFmode) probably produces incorrectly rounded results in some cases > > (though we already have such incorrect results in the other direction for > > DPD; see bug 97635). > > Thanks for the pointer. > I was aware of such imprecision in the code I introduce, but not in existing > one. I don't think this is a blocker for this patch, is it?
I don't think it's a blocker. > > Wouldn't it be better to have tests that apply for all targets supporting > > both HFmode and BID DFP, which includes x86 / x86_64 (with SSE2 support) > > as well, using _Float16 as the type name (and the float16 / > > float16_runtime effective-target keywords to test for the relevant support > > and dg-add-options float16)? > > > Indeed! What would be the appropriate directory for such tests? Maybe gcc.dg/dfp, or gcc.dg/torture if it's useful to run the tests for multiple choices of optimization options. > > That seems wrong; the __fp16 name is specific to Arm / AArch64. I'd > > expect the typedef alone to be specific; in any case, the "mode" attribute > > is the appropriate way to define this type. > > > I'll change that to __float16 (like __float128 we have a few lines below for > TFtype). There's _Float16, but no such type as __float16. -- Joseph S. Myers jos...@codesourcery.com
