Richard Henderson <richard.hender...@linaro.org> writes:
> In 88808a022c0, I tried to fix an overflow problem that affected float16 > scaling by coverting first to float64 and then rounding after that. > > However, Laurent reported that -0x3ff40000000001 converted to float16 > resulted in 0xfbfe instead of the expected 0xfbff. This is caused by > the inexact conversion to float64. > > Rather than build more logic into target/arm to compensate, just add > a function that takes a scaling parameter so that the whole thing is > done all at once with only one rounding. > > I don't have a failing test case for the float-to-int paths, but it > seemed best to apply the same solution. Can't we add the constants to the fcvt test case? > > > r~ > > > Richard Henderson (4): > softfloat: Add scaling int-to-float routines > softfloat: Add scaling float-to-int routines > target/arm: Use the int-to-float-scale softfloat routines > target/arm: Use the float-to-int-scale softfloat routines > > include/fpu/softfloat.h | 169 ++++++++---- > fpu/softfloat.c | 579 +++++++++++++++++++++++++++++++--------- > target/arm/helper.c | 130 ++++----- > 3 files changed, 628 insertions(+), 250 deletions(-) -- Alex Bennée