Re: [PING][Patch] Add support for IEEE-conformant versions of scalar fmin* and fmax*

[Richard Biener <richard.guenther@gmail.com>]

On Tue, Aug 18, 2015 at 1:07 PM, David Sherwood <david.sherwood@arm.com> wrote:
>> On Mon, Aug 17, 2015 at 11:29 AM, David Sherwood <david.sherwood@arm.com> wrote:
>> > Hi Richard,
>> >
>> > Thanks for the reply. I'd chosen to add new expressions as this seemed more
>> > consistent with the existing MAX_EXPR and MIN_EXPR tree codes. In addition it
>> > would seem to provide more opportunities for optimisation than a target-specific
>> > builtin implementation would. I accept that optimisation opportunities will
>> > be more limited for strict math compilation, but that it was still worth having
>> > them. Also, if we did map it to builtins then the scalar version would go
>> > through the optabs and the vector version would go through the target's builtin
>> > expansion, which doesn't seem very consistent.
>>
>> On another note ISTR you can't associate STRICT_MIN/MAX_EXPR and thus
>> you can't vectorize anyway?  (strict IEEE behavior is about NaNs, correct?)
> I thought for this particular case associativity wasn't an issue? We're not doing any
> reductions here, just simply performing max/min operations on each pair of elements
> in the vectors. I thought for IEEE-compliant behaviour we just need to ensure that for
> each pair of elements if one element is a NaN we return the other one.

Hmm, true.  Ok, my comment still stands - I don't see that using a tree code is
the best thing to do here.  You can add fmin/max optabs and special expansion
of BUILT_IN_FMIN/MAX and you can use a target builtin for the
vectorized variant.

The reason I am pushing against a new tree code is that we'd have an awful
lot of similar codes when pushing other flag related IL specialities to actual
IL constructs.  And we still need to find a consistent way to do that.

Richard.

> David.
>
>>
>> Richard.
>>
>> > Regards,
>> > David.
>> >
>> >> -----Original Message-----
>> >> From: Richard Biener [mailto:richard.guenther@gmail.com]
>> >> Sent: 13 August 2015 12:10
>> >> To: David Sherwood
>> >> Cc: GCC Patches
>> >> Subject: Re: [PING][Patch] Add support for IEEE-conformant versions of scalar fmin* and fmax*
>> >>
>> >> On Thu, Aug 13, 2015 at 12:11 PM, David Sherwood <david.sherwood@arm.com> wrote:
>> >> > Hi,
>> >> >
>> >> > Sorry to bother people again. Is this OK to go now?
>> >>
>> >> Hmm, why don't you go the vectorized function call path for this,
>> >> implementing the builtin_vectorized_function target hook?
>> >>
>> >> Richard.
>> >>
>> >> > Thanks!
>> >> > David.
>> >> >
>> >> >> >
>> >> >> > > On Mon, 29 Jun 2015, David Sherwood wrote:
>> >> >> > >
>> >> >> > > > Hi,
>> >> >> > > >
>> >> >> > > > I have added new STRICT_MAX_EXPR and STRICT_MIN_EXPR expressions to support the
>> >> >> > > > IEEE versions of fmin and fmax. This is done by recognising the math library
>> >> >> > > > "fmax" and "fmin" builtin functions in a similar way to how this is done for
>> >> >> > > > -ffast-math. This also allows us to vectorise the IEEE max/min functions for
>> >> >> > > > targets that support it, for example aarch64/aarch32.
>> >> >> > >
>> >> >> > > This patch is missing documentation.  You need to document the new insn
>> >> >> > > patterns in md.texi and the new tree codes in generic.texi.
>> >> >> >
>> >> >> > Hi, I've uploaded a new patch with the documentation. Hope this is ok.
>> >> >>
>> >> >> In various places where you refer to one operand being NaN, I think you
>> >> >> mean one operand being a *quiet* NaN (if one is a signaling NaN - only
>> >> >> supported by GCC if -fsignaling-nans - the IEEE minNum and maxNum
>> >> >> operations raise "invalid" and return a quiet NaN).
>> >> >
>> >> > Hi, I have a new patch that hopefully addresses the documentation issues.
>> >> >
>> >> > Thanks,
>> >> > David.
>> >> >
>> >> > ChangeLog:
>> >> >
>> >> > 2015-07-15  David Sherwood  <david.sherwood@arm.com>
>> >> >
>> >> > gcc/
>> >> >     * builtins.c (integer_valued_real_p): Add STRICT_MIN_EXPR and
>> >> >     STRICT_MAX_EXPR.
>> >> >     (fold_builtin_fmin_fmax): For strict math, convert builting fmin and
>> >> >     fmax to STRICT_MIN_EXPR and STRICT_MIN_EXPR, respectively.
>> >> >     * expr.c (expand_expr_real_2): Add STRICT_MIN_EXPR and STRICT_MAX_EXPR.
>> >> >     * fold-const.c (const_binop): Likewise.
>> >> >     (fold_binary_loc, tree_binary_nonnegative_warnv_p): Likewise.
>> >> >     (tree_binary_nonzero_warnv_p): Likewise.
>> >> >     * optabs.h (strict_minmax_support): Declare.
>> >> >     * optabs.def: Add new optabs strict_max_optab/strict_min_optab.
>> >> >     * optabs.c (optab_for_tree_code): Return new optabs for STRICT_MIN_EXPR
>> >> >     and STRICT_MAX_EXPR.
>> >> >     (strict_minmax_support): New function.
>> >> >     * real.c (real_arithmetic): Add STRICT_MIN_EXPR and STRICT_MAX_EXPR.
>> >> >     * tree.def: Likewise.
>> >> >     * tree.c (associative_tree_code, commutative_tree_code): Likewise.
>> >> >     * tree-cfg.c (verify_expr): Likewise.
>> >> >     (verify_gimple_assign_binary): Likewise.
>> >> >     * tree-inline.c (estimate_operator_cost): Likewise.
>> >> >     * tree-pretty-print.c (dump_generic_node, op_code_prio): Likewise.
>> >> >     (op_symbol_code): Likewise.
>> >> > gcc/config:
>> >> >     * aarch64/aarch64.md: New pattern.
>> >> >     * aarch64/aarch64-simd.md: Likewise.
>> >> >     * aarch64/iterators.md: New unspecs, iterators.
>> >> >     * arm/iterators.md: New iterators.
>> >> >     * arm/unspecs.md: New unspecs.
>> >> >     * arm/neon.md: New pattern.
>> >> >     * arm/vfp.md: Likewise.
>> >> > gcc/doc:
>> >> >     * generic.texi: Add STRICT_MAX_EXPR and STRICT_MIN_EXPR.
>> >> >     * md.texi: Add strict_min and strict_max patterns.
>> >> > gcc/testsuite
>> >> >     * gcc.target/aarch64/maxmin_strict.c: New test.
>> >> >     * gcc.target/arm/maxmin_strict.c: New test.
>> >
>> >
>> >
>
>
>