On Thu, Jun 5, 2025 at 7:20 PM Takayuki 'January June' Suwa
<jjsuwa_sys3...@yahoo.co.jp> wrote:
>
> On 2025/06/06 8:55, Max Filippov wrote:
>
> > On Thu, Jun 05, 2025 at 09:19:19PM +0900, Takayuki 'January June' Suwa
> > wrote:
> >> On 2025/06/05 5:09, Max Filippov wrote:
> >>> On Tue, Jun 3, 2025 at 7:44 AM Takayuki 'January June' Suwa
> >>> <jjsuwa_sys3...@yahoo.co.jp> wrote:
> >>>>
> >>>> By using the previously unused CEIL|FLOOR|ROUND.S floating-point
> >>>> coprocessor
> >>>> instructions. In addition, two instruction operand format codes are
> >>>> added
> >>>> to output the scale value as assembler source.
> >>>>
> >>>> /* example */
> >>>> int test0(float a) {
> >>>> return __builtin_lceilf(a);
> >>>> }
> >>>> int test1(float a) {
> >>>> return __builtin_lfloorf(a * 2);
> >>>> }
> >>>> int test2(float a) {
> >>>> /* __builtin_lroundf() requires -fno-math-errno */
> >>>> return __builtin_lroundf(a * 32768);
> >>>> }
> >>>>
> >>>> ;; result
> >>>> test0:
> >>>> entry sp, 32
> >>>> wfr f0, a2
> >>>> ceil.s a2, f0, 0
> >>>> retw.n
> >>>> test1:
> >>>> entry sp, 32
> >>>> wfr f0, a2
> >>>> floor.s a2, f0, 1
> >>>> retw.n
> >>>> test2:
> >>>> entry sp, 32
> >>>> wfr f0, a2
> >>>> round.s a2, f0, 15
> >>>> retw.n
> >>>>
> >>>> gcc/ChangeLog:
> >>>>
> >>>> * config/xtensa/xtensa.cc (printx, print_operand):
> >>>> Add two instruction operand format codes 'U' and 'V',
> >>>> whose represent scale factors of 0 to 15th positive/negative
> >>>> power of two.
> >>>> * gcc/config/xtensa/xtensa.md (c_enum "unspec"):
> >>>> Add UNSPEC_CEIL, UNSPEC_FLOOR and UNSPEC_ROUND.
> >>>> (int_iterator ANY_ROUND, int_attr m_round):
> >>>> New integer iterator and its attribute.
> >>>> (fix<s_fix>_truncsfsi2, *fix<s_fix>_truncsfsi2_2x,
> >>>> *fix<s_fix>_truncsfsi2_scaled, float<s_float>sisf2,
> >>>> *float<s_float>sisf2_scaled):
> >>>> Use output templates with the operand formats added above,
> >>>> instead of individual output statements.
> >>>> (l<m_round>sfsi2, *l<m_round>sfsi2_2x,
> >>>> *l<m_round>sfsi2_scaled):
> >>>> New insn patterns.
> >>>> ---
> >>>> gcc/config/xtensa/xtensa.cc | 16 ++++++++++++
> >>>> gcc/config/xtensa/xtensa.md | 52
> >>>> ++++++++++++++++++++++++++++---------
> >>>> 2 files changed, 56 insertions(+), 12 deletions(-)
> >>>
> >>> This passes without new regressions on targets without FPU, but I see
> >>> a few new failures in the gfortran testsuite on a target with FPU:
> >>>
> >>> +FAIL: gfortran.dg/nint_2.f90 -O0 execution test
> >>> +FAIL: gfortran.dg/out_of_range_1.f90 -O0 execution test
> >>> +FAIL: gfortran.dg/out_of_range_1.f90 -O1 execution test
> >>> +FAIL: gfortran.dg/out_of_range_1.f90 -O2 execution test
> >>> +FAIL: gfortran.dg/out_of_range_1.f90 -O3 -fomit-frame-pointer
> >>> -funroll-loops -fpeel-loops -ftracer -finline-functions execution
> >>> test
> >>> +FAIL: gfortran.dg/out_of_range_1.f90 -O3 -g execution test
> >>> +FAIL: gfortran.dg/out_of_range_1.f90 -Os execution test
> >>> +FAIL: gfortran.dg/out_of_range_2.f90 -O0 execution test
> >>> +FAIL: gfortran.dg/out_of_range_2.f90 -O1 execution test
> >>> +FAIL: gfortran.dg/out_of_range_2.f90 -O2 execution test
> >>> +FAIL: gfortran.dg/out_of_range_2.f90 -O3 -fomit-frame-pointer
> >>> -funroll-loops -fpeel-loops -ftracer -finline-functions execution
> >>> test
> >>> +FAIL: gfortran.dg/out_of_range_2.f90 -O3 -g execution test
> >>> +FAIL: gfortran.dg/out_of_range_2.f90 -Os execution test
> >>>
> >>> At first glance they may be related to the rounding mode settings,
> >>> let me take a closer look.
> >>>
> >> gccint says l(ceil/floor/round)sfsi2 are not interested in the current
> >> rounding mode.
> >
> > I couldn't find anything about how lroundsfsi2 should behave in the
> > gccint, was looking for how exactly the rounding is supposed to happen.
> >
>
> 'lroundmn2'
> Convert operand 1 (valid for floating point mode m) to fixed point
> mode n as a signed number rounding to nearest and away from zero and
> ^^^^^^^^^^^^^^
> store in operand 0 (which has mode n).
>
> -- 16.10 Standard Pattern Names For Generation, gccint
>
> >> Similary, Xtensa ISA refman says CEIL/FLOOR/ROUND.S will
> >> not affect the current rounding mode.
> >>
> >> IMHO I suspect that the behavior for out-of-range input values, e.g. values
> >> that do not fit in signed 32 bits, infinity or NaN, may be different.
> >
> > The issue is that round.s opcode rounds to the nearest multiple of 2,
> > not away from zero as expected by the failing tests, AFAICT.
> > So e.g. -128.5 is rounded by round.s to -128.
> > I observe this with both FPU2000 and DFPU Xtensa options.
>
> ISA refman does not mention how to round half but according to your
> observation, ROUND.S will round half toward even...
>
> Ah, OK, I retract the patch.
I'd just drop the lround part. The other two look good.
If you agree I'll commit your patch with that change.
--
Thanks.
-- Max
