[PING][Patch 3/3][Arm] Add support for IEEE-conformant versions of scalar fmin* and fmax*

[David Sherwood]

Hi Kyrill,

Thanks for the reply, I think this latest patch addresses your
comments. I have added a test for the scalar forms of fmax/fmin.

Is this ok now?

Tested:

arm-none-eabi: no regressions

Cheers,
Dave.

ChangeLog:

2015-12-22  David Sherwood  <david.sherwood@arm.com>

    gcc/
        * config/arm/iterators.md (VMAXMINFNM): New int iterator.
        (fmaxmin): New int attribute.
        (fmaxmin_op): Likewise.
        * config/arm/unspecs.md (UNSPEC_VMAXNM): New unspec.
        (UNSPEC_VMINNM): Likewise.
        * config/arm/neon.md (fmaxmin): New pattern.
        * config/arm/vfp.md (fmaxmin): Likewise.
    gcc/testsuite
        * gcc.target/arm/fmaxmin.x: New file used by tests below.
        * gcc.target/arm/fmaxmin.c: New test.
        * gcc.target/arm/vect-fmaxmin.c: Likewise.



________________________________________
From: Kyrill Tkachov <kyrylo.tkachov@foss.arm.com>
Sent: Wednesday, December 16, 2015 9:31 AM
To: David Sherwood; GCC Patches
Subject: Re: [PING][Patch 3/3][Arm] Add support for IEEE-conformant versions of scalar fmin* and fmax*

Hi David,

On 16/12/15 08:53, David Sherwood wrote:
> Hi,
>
> Here is the last patch of the fmin/fmax change, which adds the optabs
> to the arm backend.
>
> Tested:
>
> arm-none-eabi: no regressions
>
> Good to go?
> David Sherwood.
>
> ChangeLog:
>
> 2015-12-08  David Sherwood  <david.sherwood@arm.com>
>
>      gcc/
>          * config/arm/iterators.md: New iterators.
>          * config/arm/unspecs.md: New unspecs.
>          * config/arm/neon.md: New pattern.
>          * config/arm/vfp.md: Likewise.

Please list the new entities you add in parentheses.
For example:
     * config/arm/iterators.md (VMAXMINFNM): New int iterator.
     (fmaxmin): New int attribute.
     (fmaxmin): Likewise.

Same for the other files. That way one can grep through the ChangeLogs to
find when any particular pattern/iterator/whatever was modified.

+;; Auto-vectorized forms for the IEEE-754 fmax()/fmin() functions
+(define_insn "<fmaxmin><mode>3"
+  [(set (match_operand:VCVTF 0 "s_register_operand" "=w")
+       (unspec:VCVTF [(match_operand:VCVTF 1 "s_register_operand" "w")
+                      (match_operand:VCVTF 2 "s_register_operand" "w")]
+                      VMAXMINFNM))]
+  "TARGET_NEON && TARGET_FPU_ARMV8"
+  "<fmaxmin_op>.<V_s_elem>\t%<V_reg>0, %<V_reg>1, %<V_reg>2"
+  [(set_attr "type" "neon_fp_minmax_s<q>")]
+)

I would just say "Vector forms" rather than "Auto-vectorized".
In principle we can get vector types through other means besides
auto-vectorisation.

+;; Scalar forms for the IEEE-754 fmax()/fmin() functions
+(define_insn "<fmaxmin><mode>3"
+  [(set (match_operand:SDF 0 "s_register_operand" "=<F_constraint>")
+       (unspec:SDF [(match_operand:SDF 1 "s_register_operand" "<F_constraint>")
+                    (match_operand:SDF 2 "s_register_operand" "<F_constraint>")]
+                    VMAXMINFNM))]
+  "TARGET_HARD_FLOAT && TARGET_VFP5 <vfp_double_cond>"
+  "<fmaxmin_op>.<V_if_elem>\\t%<V_reg>0, %<V_reg>1, %<V_reg>2"
+  [(set_attr "type" "f_minmax<vfp_type>")
+   (set_attr "conds" "unconditional")]
+)
+

I notice your new test doesn't test the SF variant of this.
Could you please add something to test it?
Looks good to me otherwise.

Thanks,
Kyrill

-------------- next part --------------
A non-text attachment was scrubbed...
Name: fmaxmin3.patch
Type: text/x-patch
Size: 6127 bytes
Desc: fmaxmin3.patch
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20151222/3f0347eb/attachment.bin>