Re: [PATCH] ppc: Fix 'calli' when floating-point arguments are passed

[Paulo César Pereira de Andrade]

Em ter., 23 de ago. de 2022 às 13:09, Paul Cercueil
<paul@crapouillou.net> escreveu:
>
> Actually I think my calli() fix is wrong...
>
> I swear the 2to3 check was passing before (with the exact same patch I
> sent) and now it decided to fail again.
>
> I think it's because I added an instruction before the BL(d), so the
> value of "d" is outdated. The calculation of "d" should probably be:
> d = (i0 - _jit->pc.w - !!varargs * 4) & ~3;
>
> With this changed the 2to3 check works again here.

  Fixed.

> As for the SQRT, with the following simplified test:
>
> ---
> #include "alu.inc"
>
> .code
>  prolog
> #define SQRT(N, T, I, V) FUN1(N, T, sqrt, I, V, f0, f1)
>   SQRT(0, _f, -0.0, 0.0)
>
>  prepare
>   pushargi ok
>   ellipsis
>  finishi @printf
>  ret
>  epilog
> ---
>
> I get:
>
> paul@frank:~/dev/lightning$ check/lightning -v check/fop_sqrt.tst
>         0x3f680000 mflr r0
>         0x3f680004 stw r0,4(r1)
>         0x3f680008 stw r14,-72(r1)
>         0x3f68000c stfd f14,-80(r1)
>         0x3f680010 stfd f15,-88(r1)
>         0x3f680014 stw r31,-4(r1)
>         0x3f680018 mr r31,r1
>         0x3f68001c stwu r1,-160(r1)
>         0x3f680020 lis r14,16232
>         0x3f680024 lfs f15,4116(r14)
>         0x3f680028 fsqrts f14,f15
>         0x3f68002c lis r14,16232
>         0x3f680030 lfs f0,4120(r14)
>         0x3f680034 fcmpo cr0,f14,f0
>         0x3f680038 beq 0x3f680040 :check/fop_sqrt.tst:7
>         0x3f68003c bl 0x3f2b75f0
> # :check/fop_sqrt.tst:7
>         0x3f680040 lis r14,16232
>         0x3f680044 lfs f14,4116(r14)
>         0x3f680048 fsqrts f14,f14
>         0x3f68004c lis r14,16232
>         0x3f680050 lfs f0,4120(r14)
>         0x3f680054 fcmpo cr0,f14,f0
>         0x3f680058 beq 0x3f680060 :check/fop_sqrt.tst:7
>         0x3f68005c bl 0x3f2b75f0
>         0x3f680060 lis r14,16232
>         0x3f680064 lfs f14,4116(r14)
>         0x3f680068 fsqrts f15,f14
>         0x3f68006c lis r14,16232
>         0x3f680070 lfs f0,4120(r14)
>         0x3f680074 fcmpo cr0,f15,f0
>         0x3f680078 beq 0x3f680080 :check/fop_sqrt.tst:7
>         0x3f68007c bl 0x3f2b75f0
>         0x3f680080 lis r14,16232
>         0x3f680084 lfs f15,4116(r14)
>         0x3f680088 fsqrts f15,f15
>         0x3f68008c lis r14,16232
>         0x3f680090 lfs f0,4120(r14)
>         0x3f680094 fcmpo cr0,f15,f0
>         0x3f680098 beq 0x3f6800a0 :check/fop_sqrt.tst:7
>         0x3f68009c bl 0x3f2b75f0
>         0x3f6800a0 lis r3,16388
>         0x3f6800a4 ori r3,r3,20992
>         0x3f6800a8 bl 0x3f2ee17c
>         0x3f6800ac addi r1,r1,160
>         0x3f6800b0 lwz r0,4(r1)
>         0x3f6800b4 lwz r14,-72(r1)
>         0x3f6800b8 lfd f14,-80(r1)
>         0x3f6800bc lfd f15,-88(r1)
>         0x3f6800c0 mtlr r0
>         0x3f6800c4 lwz r31,-4(r1)
>         0x3f6800c8 blr
>   - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
> - - -
> qemu: uncaught target signal 4 (Illegal instruction) - core dumped

https://www.ibm.com/docs/en/aix/7.1?topic=set-fsqrts-floating-square-root-single-instruction
says:
"""
Implementation

This instruction is optionally defined for PowerPC® implementations.
Using it on an implementation that does not support this instruction
will cause the system illegal instruction error handler to be invoked.

This instruction is an optional instruction of the PowerPC®
architecture and may not be implemented in all machines.
"""
  Can you see what /proc/cpuinfo says? Maybe if we know it is
not available, do a FMR, FSQRT, FRSP (extr_f_d, sqrtr_d, extr_d_f).

> Cheers,
> -Paul
>
>
>
> Le mar., août 23 2022 at 12:45:51 -0300, Paulo César Pereira de
> Andrade <paulo.cesar.pereira.de.andrade@gmail.com> a écrit :
> > Em ter., 23 de ago. de 2022 às 12:31, Paul Cercueil
> > <paul@crapouillou.net> escreveu:
> >>
> >>  Note that this patch fixes the 3to2 and varargs checks, but the
> >>  following are still failing: cvt, fop_sqrt, clobber, call.
> >
> >   Let's first check the most likely cause. For fop_sqrt.tst, please
> > test commenting with // or /* */ the lines that have either $NaN
> > or $Inf.
> >   Also in the first suspect, clobber.tst might be failing at:
> >          jmpf(__LINE__, unord,       0, $(0.0 / 0.0))
> > but please test bisecting the tests, until finding the one that fails.
> > It must be calling abort, running in a debugger would be a bit
> > easier to debug.
> >
> >   For call.tst, please also bisect to find what test is failing. This
> > test does extensive checking on to/from float/double tests, and
> > to/from float/double/word tests as well as several validations on
> > special values, for 32 bit, values like 0x80000000, 0xffffffff,
> > 0x7fffffff, 0x80000001 and the similar 8 and 16 bit values, to
> > check for proper conversions.
> >
> >   It must be something simple like the patch you just did, or
> > might be a qemu-user bug.
> >
> >>  Cheers,
> >>  -Paul
> >
> > Thanks,
> > Paulo
>
>