[PATCH 3/4] Add ILP32 support to aarch64

[Steve Ellcey]

On Tue, 2017-08-08 at 16:01 +0100, Szabolcs Nagy wrote:
>Â 
> > +#if IREG_SIZE == 64 && OREG_SIZE == 32
> > +Â Â if (__builtin_fabs (x) > INT32_MAX - 2)
> i don't understand the -2 here.

I was confused and trying to handle the fact that fabs(INT32_MIN) !=
INT32_MAX. Â I have removed the -2 and am just comparing to INT32_MAX
and that seems to work fine. Â Since fabs(INT32_MIN) is greater than
INT32_MAX we may unnecessarily enter this if statement for values
between  INT32_MIN and INT32_MIN+1 but that should not cause any
failures, just a slowdown.

> > +Â Â Â Â {
> > +Â Â Â Â Â Â /* Converting large values to a 32 bit in may cause the
> > frintx/fcvtza
> s/in/int/

Fixed that.

> > +Â Â Â Â Â Â invalid_p = libc_fetestexcept (FE_INVALID);
> > +Â Â Â Â Â Â inexact_p = libc_fetestexcept (FE_INEXACT);
> multiple flags can be tested/raised in a single call.

Good point. Â I changed this to one call and saved the flags in an
integer variable for checking later.

> > +Â Â Â Â Â Â libc_fesetenv (&env);
> > +
> > +Â Â Â Â Â Â if (invalid_p)
> > +	feraiseexcept (FE_INVALID);
> > +Â Â Â Â Â Â else if (inexact_p)
> > +	feraiseexcept (FE_INEXACT);
> > +
> i think correct trapping is not guaranteed by glibc,
> only correct status flags when the function returns,
> so spurious inexact is not a problem if it is already
> raised, and then i expect better code gen for the
> inexact clearing approach:
> 
> if (fabs (x) > INT32_MAX && fetestexcept (FE_INEXACT) == 0)
> Â  {
> Â Â Â Â asm (...);
> Â Â Â Â if (fetestexcept (FE_INVALID|FE_INEXACT) ==
> (FE_INVALID|FE_INEXACT))
> Â Â Â Â Â Â feclearexcept (FE_INEXACT);
> Â  }
> else
> Â  asm (...);

As you mentioned in your followup email, we have to worry about
FE_INVALID being set on entry too. Â I have attached an updated
version of my patch.

Steve Ellcey
sellcey@cavium.com


2017-08-08  Steve Ellcey  <sellcey@cavium.com>

	* sysdeps/aarch64/fpu/s_llrint.c (OREG_SIZE): New macro.
	* sysdeps/aarch64/fpu/s_llround.c (OREG_SIZE): Likewise.
	* sysdeps/aarch64/fpu/s_llrintf.c (OREGS, IREGS): Remove.
	(IREG_SIZE, OREG_SIZE): New macros.
	* sysdeps/aarch64/fpu/s_llroundf.c: (OREGS, IREGS): Remove.
	(IREG_SIZE, OREG_SIZE): New macros.
	* sysdeps/aarch64/fpu/s_lrintf.c (IREGS): Remove.
	(IREG_SIZE): New macro.
	* sysdeps/aarch64/fpu/s_lroundf.c (IREGS): Remove.
	(IREG_SIZE): New macro.
	* sysdeps/aarch64/fpu/s_lrint.c (math_private.h, fenv.h, stdint.h):
	New includes.
	(IREG_SIZE, OREG_SIZE): Initialize if not already set.
	(OREGS, IREGS): Set based on IREG_SIZE and OREG_SIZE.
	(__CONCATX): Handle exceptions correctly on large values that may
	set FE_INVALID.
	* sysdeps/aarch64/fpu/s_lround.c (IREG_SIZE, OREG_SIZE):
	Initialize if not already set.
        (OREGS, IREGS): Set based on IREG_SIZE and OREG_SIZE.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: math.patch
Type: text/x-patch
Size: 4844 bytes
Desc: not available
URL: <http://sourceware.org/pipermail/libc-alpha/attachments/20170808/2a9f471f/attachment.bin>