| Message ID | 20200706234737.32378-4-jcmvbkbc@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | target/xtensa: implement double precision FPU | expand |
Max Filippov <jcmvbkbc@gmail.com> writes: > pickNaNMulAdd logic on Xtensa is the same as pickNaN when applied to > the expression (a * b) + c. So with two pickNaN variants there must be > two pickNaNMulAdd variants. > > Cc: Peter Maydell <peter.maydell@linaro.org> > Cc: "Alex Bennée" <alex.bennee@linaro.org> > Signed-off-by: Max Filippov <jcmvbkbc@gmail.com> Acked-by: Alex Bennée <alex.bennee@linaro.org>
On 7/6/20 4:47 PM, Max Filippov wrote: > pickNaNMulAdd logic on Xtensa is the same as pickNaN when applied to > the expression (a * b) + c. So with two pickNaN variants there must be > two pickNaNMulAdd variants. "Is the same as"? I question the non-use of the infzero parameter. When infzero, (a * b) = (Inf * 0), which will produce a default QNaN. Your sentence above would suggest that pickNaN is applied twice, so that if use_first_nan, the default nan is chosen above any nan within c. In addition, is the invalid flag raised for (Inf * 0) + NaN? Does that happen regardless of the use_first_nan setting, or does the whole operation short-circuit? r~
On Wed, Jul 8, 2020 at 9:07 AM Richard Henderson <richard.henderson@linaro.org> wrote: > > On 7/6/20 4:47 PM, Max Filippov wrote: > > pickNaNMulAdd logic on Xtensa is the same as pickNaN when applied to > > the expression (a * b) + c. So with two pickNaN variants there must be > > two pickNaNMulAdd variants. > > "Is the same as"? > > I question the non-use of the infzero parameter. > > When infzero, (a * b) = (Inf * 0), which will produce a default QNaN. Your > sentence above would suggest that pickNaN is applied twice, so that if > use_first_nan, the default nan is chosen above any nan within c. Apparently the description it's wrong: only NaNs in arguments are considered, default NaN produced as a result of (a * b) is never chosen when c is NaN. > In addition, is the invalid flag raised for (Inf * 0) + NaN? Does that happen > regardless of the use_first_nan setting, or does the whole operation short-circuit? Flag setting happens regardless of which NaN is chosen. I'll post v3 with a fix for this and additional tests.
diff --git a/fpu/softfloat-specialize.inc.c b/fpu/softfloat-specialize.inc.c index e17327b405c5..6b07d50efad3 100644 --- a/fpu/softfloat-specialize.inc.c +++ b/fpu/softfloat-specialize.inc.c @@ -574,6 +574,24 @@ static int pickNaNMulAdd(FloatClass a_cls, FloatClass b_cls, FloatClass c_cls, } else { return 1; } +#elif defined(TARGET_XTENSA) + if (status->use_first_nan) { + if (is_nan(a_cls)) { + return 0; + } else if (is_nan(b_cls)) { + return 1; + } else { + return 2; + } + } else { + if (is_nan(c_cls)) { + return 2; + } else if (is_nan(b_cls)) { + return 1; + } else { + return 0; + } + } #else /* A default implementation: prefer a to b to c. * This is unlikely to actually match any real implementation.
pickNaNMulAdd logic on Xtensa is the same as pickNaN when applied to the expression (a * b) + c. So with two pickNaN variants there must be two pickNaNMulAdd variants. Cc: Peter Maydell <peter.maydell@linaro.org> Cc: "Alex Bennée" <alex.bennee@linaro.org> Signed-off-by: Max Filippov <jcmvbkbc@gmail.com> --- fpu/softfloat-specialize.inc.c | 18 ++++++++++++++++++ 1 file changed, 18 insertions(+)