MIPS: IEEE 754-2008 features support

On Tue, 16 Jul 2013, Richard Sandiford wrote:

> >  The run-time selection options are -mnan=2008 and -mnan=legacy for the 
> > 2008 NaN and the legacy NaN mode respectively.  The binary patterns of NaN 
> > data produced by compiler builtins or otherwise is adjusted accordingly, 
> > as is a `.nan' directive included now with the assembly source produced, 
> > used by GAS to set the NAN2008 ELF file header flag appropriately.  These 
> > options are also passed down to GAS by the GCC driver so that user 
> > assembly sources have the said NAN2008 flag set correctly in the absence 
> > of a `.nan' directive.  Finally the name of the dynamic linker is changed 
> > accordingly if the driver is used to invoke the linker.
> 
> I was a bit surprised that a change to the NaN format requires a change
> to the dynamic linker, but I suppose that's the consequence of forcing
> every ELF object to fall on one side of the fence?

 Yes, the dynamic linker has to enforce NAN2008 ELF file header flag 
compatibility among the modules loaded.  The file name has to change to 
avoid accidentally loading an executable with a pre-NAN2008 dynamic linker 
that wouldn't enforce that requirement.

> I assume eventually there'll be an -mmac={legacy|2008} option too,
> although of course it's good that that's a separate patch.  But then
> the question is: is there a plan for handling all 8 possible combinations
> of -mnan, -mabs and -mmac?  How would the FCSR be set up?

 The NAN2008 FCSR bit is supposed to be preset by startup code (for bare 
iron) or the kernel (for Linux) according to the NAN2008 ELF file header 
flag.  That code is also responsible for checking that the processor 
supports the required FPU configuration (i.e. the NAN2008 FCSR bit is 
either writable or hardwired to the required setting) and e.g. returning 
from the execve(2) call if not.

 I have no information on any such arrangement for the ABS2008 bit or any 
plans on a -mmac= option at this time.  Please note that FCSR is 
user-writable (requires no CP0 privilege) anyway, so ABS2008 can be set in 
main or wherever according to one's requirements.

> Even if we do plan to have all three be separately configurable,
> I imagine most people will want them all set the same way, so it would
> probably make sense to have something like -mfp-model={legacy|2008}
> for that case.  And then my inclination would be to start with that
> option and only add the split-out versions if we have a specific
> (rather than theoretical) use case for them.

 This is risky because of the ABS2008 setting implication noted above.  
The current situation is a 2008-NaN executable will be started with the 
ABS2008 FCSR bit clear (unless hardwired to 1 in the processor, that is).

 Please also note that the writability of the individual new (HAS2008) 
FCSR bits is optional e.g. a conforming processor may have NAN2008 
hardwired to 1 and ABS2008 hardwired to 0 (or likewise with NAN2008 
writable).

> The patch mostly looks good apart from that, but please use a single
> enum for the 2008/legacy thing, both in mips.h and mips.opt.

 Also mips.c and mips.md (and last but not least mips-opts.h).  Done.

> > +;; On pre IEEE 754-2008 processors abs.fmt is an arithmetic instruction
> > +;; and treats all NaN inputs as invalid; it does not clear their sign bit.
> > +;; We therefore can't use abs.fmt unless in the IEEE 754-2008 mode or if
> > +;; the signs of NaNs matter.
> 
> "don't matter".  Or maybe to avoid the double negative:
> 
> ;; The "legacy" (as opposed to "2008") form of ABS.FMT is an arithmetic
> ;; instruction that treats all NaN inputs as invalid; it does not flip
> ;; their sign bit.  We therefore can't use that form if the signs of
> ;; NaNs matter.

 That sounds better, thanks.  Since you suggested taking the opportunity 
and changing the capitalisation of the mnemonic I've decided to use 
ABS.fmt/NEG.fmt instead per the usual convention (`fmt' is a variable 
part).  I also used "clear" rather than "flip" for ABS.fmt as in the 
original version.

> > +These options control the treatment of the special not-a-number (NaN)
> > +IEEE 754 floating-point data with the @code{abs.@i{fmt}} and
> > +@code{neg.@i{fmt}} machine instructions.
> > +
> > +By default or when the @option{-mabs=legacy} is used the legacy
> > +treatment is selected.  In this case these instructions are considered
> > +arithmetic and avoided where correct operation is required and the
> > +input operand might be a NaN.  A longer sequence of instructions that
> > +manipulate the sign bit of floating-point datum manually is used
> > +instead unless the @option{-ffinite-math-only} option has also been
> > +specified.
> > +
> > +The @option{-mabs=2008} option selects the IEEE 754-2008 treatment.  In
> 
> I think it reads better without the "-2008" bit, since even the 1985
> version suggested this behaviour in the appendix.

 Except the appendix was clearly marked as not a part of the standard 
itself, and moreover that was merely a suggestion.  Now the semantics is a 
part of the standard proper (Section 5.5.1 "Sign bit operations") and is a 
hard requirement: "The operations treat floating-point numbers and NaNs 
alike, and signal no exception," so I'd prefer to keep it.

 If you still think my chosen wording is unfortunate for some reason, then 
please propose an alternative which avoids the ambiguity as to the version 
of the standard.

> > +The @option{-mnan=legacy} option selects the legacy encoding.  In this
> > +case quiet NaNs (qNaNs) are denoted by the first bit of their trailing
> > +significand field being 0, whereas signalling NaNs (sNaNs) are denoted
> > +by the first bit of their trailing significand field being 1.
> > +
> > +The @option{-mnan=2008} option selects the IEEE 754-2008 encoding.  In
> > +this case qNaNs are denoted by the first bit of their trailing
> > +significand field being 1, whereas sNaNs are denoted by the first bit of
> > +their trailing significand field being 0.
> 
> Very minor, but I'd prefer this without the "trailing"s.

 Are you sure?  Please note that "trailing significand field" (another 
standard's term, just like "symbolic datum" discussed previously) 
emphasises the leading bit of the significand being implied rather than 
encoded in the field.

> > +These options are only supported on targets using the ELF binary format.
> 
> Let's drop that bit.  We don't support non-ELF on MIPS.

 I wasn't sure, removed now.  When was non-ELF support dropped?  I used to 
use GCC (or more precisely the EGCS fork) on DEC Ultrix.

> > +/* { dg-do compile } */
> > +/* { dg-options "-mhard-float -mabs=2008" } */
> 
> Please instead make "-mabs=" imply "-mhard-float" in mips.exp.
> (It is an "imply"; you can still test for -msoft-float -mabs=...
> if you want.)

 Like this?

> > +# Return 1 if this is a MIPS target supporting -mnan=.
> > +# Old versions of binutils may not support this option.
> > +
> > +proc check_effective_target_mips_nan { } {
> > +    if { ![istarget mips*-*-*] } {
> > +	return 0
> > +    }
> > +    return [check_no_compiler_messages mips_nan object {
> > +	int dummy;
> > +    } "-mnan=2008"]
> > +}
> 
> The tests you added are dg-do compile tests, which stop after assembly
> generation, so this guard shouldn't be needed.

 It is needed in case the compiler was built without support for this 
option i.e. configured against old binutils.  I verified that it indeed 
triggered in this case:

Executing on host: mips-linux-gnu-gcc  -fno-diagnostics-show-caret -fdiagnostics-color=never  -mnan=2008 -c  -o mips_nan21474.o mips_nan21474.c    (timeout = 300)
mips-linux-gnu-gcc: error: unrecognized command line option '-mnan=2008'
compiler exited with status 1
output is:
mips-linux-gnu-gcc: error: unrecognized command line option '-mnan=2008'

UNSUPPORTED: gcc.target/mips/nan-2008.c  -O0
UNSUPPORTED: gcc.target/mips/nan-2008.c  -O1

-- and so on.

>  FWIW, if you did have
> dg-do assemble or dg-do run tests, the mips.exp approach would be to
> downgrade them to dg-do compile rather than skip them altogether.

 Except that the compiler would still lack support for -mnan=2008.

  Maciej

2013-07-17  Maciej W. Rozycki  <macro@codesourcery.com>

	gcc/
	* config/mips/linux.h (GLIBC_DYNAMIC_LINKER): Handle `-mnan=2008'.
	(UCLIBC_DYNAMIC_LINKER): New macro.
	* config/mips/linux64.h (GLIBC_DYNAMIC_LINKER32): Handle 
	`-mnan=2008'.
	(GLIBC_DYNAMIC_LINKER64, GLIBC_DYNAMIC_LINKERN32): Likewise.
	(UCLIBC_DYNAMIC_LINKER32): Undefine macro first.  Handle 
	`-mnan=2008'.
	(UCLIBC_DYNAMIC_LINKER64): Redefine macro.
	(UCLIBC_DYNAMIC_LINKERN32): Likewise.
	* config/mips/mips-modes.def: Remove RESET_FLOAT_FORMAT calls
	for SF and DF modes.  Use ieee_quad_format for TF mode.
	* config/mips/mips-opts.h (mips_ieee_754_setting): New enum.
	* config/mips/mips.c (mips_file_start) [HAVE_AS_NAN]: Output a 
	`.nan' directive.
	(mips_option_override): Handle `-mnan=legacy'.
	* config/mips/mips.h (TARGET_CPU_CPP_BUILTINS): Handle 
	`-mabs=2008' and `-mnan=2008'.
	(OPTION_DEFAULT_SPECS): Add "nan" default.
	(ASM_SPEC): Handle `-mnan='.
	* config/mips/mips.md (abs<mode>2): Handle `-mabs=2008', update
	comment accordingly.
	(neg<mode>2): Likewise.
	* config/mips/mips.opt (mabs, mnan): New options.
	* doc/install.texi (Configuration): Document `--with-nan=' option.
	* doc/invoke.texi (Option Summary): List MIPS `-mabs=' and 
	`-mnan=' options.
	(MIPS Options): Document them.
	* config.gcc <mips*-*-*>: Handle `--with-nan='.
	* configure.ac <mips*-*-*>: Check for GAS `-mnan=2008' support.
	* configure: Regenerate.
	* config.in: Regenerate.

	gcc/testsuite/
	* gcc.target/mips/fabs-2008.c: New test case.
	* gcc.target/mips/fabs-legacy.c: New test case.
	* gcc.target/mips/fabsf-2008.c: New test case.
	* gcc.target/mips/fabsf-legacy.c: New test case.
	* gcc.target/mips/fneg-2008.c: New test case.
	* gcc.target/mips/fneg-legacy.c: New test case.
	* gcc.target/mips/fneg-2008.c: New test case.
	* gcc.target/mips/fneg-legacy.c: New test case.
	* gcc.target/mips/nan-2008.c: New test case.
	* gcc.target/mips/nan-legacy.c: New test case.
	* gcc.target/mips/nanf-2008.c: New test case.
	* gcc.target/mips/nanf-legacy.c: New test case.
	* gcc.target/mips/nans-2008.c: New test case.
	* gcc.target/mips/nans-legacy.c: New test case.
	* gcc.target/mips/nansf-2008.c: New test case.
	* gcc.target/mips/nansf-legacy.c: New test case.
	* gcc.target/mips/mips.exp: Handle `-mabs=' and `-mnan='.
	* lib/target-supports.exp (check_effective_target_mips_nan): New 
	procedure.

  Maciej

gcc-mips-nan2008.diff

MIPS: IEEE 754-2008 features support

Commit Message

Comments

Patch