Message ID | 20220328201442.175206-1-matheus.ferst@eldorado.org.br |
---|---|
Headers | show |
Series | softfloat 128-bit integer support | expand |
On 3/28/22 14:14, matheus.ferst@eldorado.org.br wrote: > From: Matheus Ferst <matheus.ferst@eldorado.org.br> > > This RFC is a first attempt at implementing the 128-bit integer > conversion routines in softfloat, as required by the xscv[su]qqp and > xscvqp[su]qz instructions of PowerISA v3.1. > > Instead of using int128.h, int-to-float routines receive the 128-bit > numbers through a pair of 64-bit values, and float-to-int conversions > use a pointer to return the lower half of the result. > > We only need the parts128 methods, but since the difference to parts64 > ones seemed minor, I included both in this patch. > > RFC: > - Should we use struct Int128 instead of 64-bit value pairs? I think so. We have it, and it makes the interface more obvious. > - I've not tested the float64 methods since the PPC instructions only > use the quad-precision routines. Should we keep them in the final > version? Let's not add anything that we don't have a need for. It may eventually be needed by RISC-V RV128, but we can add it then. r~
On 29/03/2022 00:38, Richard Henderson wrote: > On 3/28/22 14:14, matheus.ferst@eldorado.org.br wrote: >> From: Matheus Ferst <matheus.ferst@eldorado.org.br> >> >> This RFC is a first attempt at implementing the 128-bit integer >> conversion routines in softfloat, as required by the xscv[su]qqp and >> xscvqp[su]qz instructions of PowerISA v3.1. >> >> Instead of using int128.h, int-to-float routines receive the 128-bit >> numbers through a pair of 64-bit values, and float-to-int conversions >> use a pointer to return the lower half of the result. >> >> We only need the parts128 methods, but since the difference to parts64 >> ones seemed minor, I included both in this patch. >> >> RFC: >> - Should we use struct Int128 instead of 64-bit value pairs? > > I think so. We have it, and it makes the interface more obvious. > >> - I've not tested the float64 methods since the PPC instructions only >> use the quad-precision routines. Should we keep them in the final >> version? > > Let's not add anything that we don't have a need for. > It may eventually be needed by RISC-V RV128, but we can add it then. > > > r~ Thanks for your comments and review. I'll send an alternative version of this RFC using Int128.
From: Matheus Ferst <matheus.ferst@eldorado.org.br> This RFC is a first attempt at implementing the 128-bit integer conversion routines in softfloat, as required by the xscv[su]qqp and xscvqp[su]qz instructions of PowerISA v3.1. Instead of using int128.h, int-to-float routines receive the 128-bit numbers through a pair of 64-bit values, and float-to-int conversions use a pointer to return the lower half of the result. We only need the parts128 methods, but since the difference to parts64 ones seemed minor, I included both in this patch. RFC: - Should we use struct Int128 instead of 64-bit value pairs? - I've not tested the float64 methods since the PPC instructions only use the quad-precision routines. Should we keep them in the final version? Matheus Ferst (6): softfloat: add uint128_to_float* conversion methods softfloat: add int128_to_float* conversion methods softfloat: add float*_to_uint128 conversion methods softfloat: add float*_to_int128 conversion methods target/ppc: implement xscv[su]qqp target/ppc: implement xscvqp[su]qz fpu/softfloat-parts.c.inc | 202 ++++++++++++++++++++++++++++ fpu/softfloat.c | 161 ++++++++++++++++++++++ include/fpu/softfloat.h | 23 ++++ target/ppc/fpu_helper.c | 34 +++++ target/ppc/helper.h | 4 + target/ppc/insn32.decode | 7 + target/ppc/translate/vsx-impl.c.inc | 22 +++ 7 files changed, 453 insertions(+)