Hi! On Fri, Apr 07, 2023 at 02:34:01AM -0400, Michael Meissner wrote: > As we discussed in a private chat room, I modified the code to generate > vmaddfp > and vnmsubfp if -Ofast (-ffast-math) is used.
As I said, that is no good. > This allows the compiler to > eliminate the extra move if the user does not care about strict floating point > code generation, but it generates only the VSX instructions in the normal > case. You should not generate *any* VMX computational insns unless the user asked for that *explicitly*. Not only the rounding mode matters (always RN=00 for VMX), but also the NJ setting, and the default for NJ is unusable for normal code (that is, code that is not low-precision graphics code or the like; most code). Please change *only* the two patterns I mentioned? Just never generate vmaddfp or vnmsubfp when not explicitly asked for it. Segher
